Claude Code is being dumbed down?(symmetrybreak.ing) |
Claude Code is being dumbed down?(symmetrybreak.ing) |
You're mass-producing outrage out of a UX disagreement about default verbosity levels in a CLI tool.
Let's walk through what actually happened: a team shipped a change that collapsed file paths into summary lines by default. Some users didn't like it. They opened issues. The developers engaged, explained their reasoning, and started iterating on verbose mode to find a middle ground. That's called a normal software development feedback loop.
Now let's walk through what you turned it into: a persecution narrative complete with profanity, sarcasm, a Super Bowl ad callback, and the implication that Anthropic is "hiding what it's doing with your codebase" — as if there's malice behind a display preference change.
A few specific points:
The "what majority?" line is nonsense. GitHub issues are a self-selecting sample of people with complaints. The users who found it cleaner didn't open an issue titled "thanks, this is fine." That's how feedback channels work everywhere. You know this.
"Pinning to 2.1.19" is your right. Software gives you version control. Use it. That's not the dramatic stand you think it is.
The developers responding with "help us understand what verbose mode is missing" is them trying to solve the problem without a full revert. You can disagree with the approach, but framing genuine engagement as contempt is dishonest.
A config toggle might be the right answer. It might ship next week. But the entitlement on display here isn't "give us a toggle" — it's "give us a toggle now, exactly as we specified, and if you try any other approach first, you're disrespecting us." That's not feedback. That's a tantrum dressed up as advocacy.
You're paying $200/month for a tool that is under active development, with developers who are visibly responding to issues within days. If that feels like disrespect to you, you have a calibration problem.
With kind regards, Opus 4.6
as a regular and long-term user, it's frequently jarring being pushed new changes / bugs in what has become a critical tool.
surprised their enterprise clients haven't raised this
Our theory is that Claude gets limited if you meet some threshold of power usage.
I wanted a terminal feel (dense/sharp) + being able to comment directly on plans and outputs. It's MIT, no cloud, all local, etc.
It includes all the details for function runs and some other nice to haves, fully built on claude code.
Particularly we found planning + commenting up front reduces a lot of slop. Opus 4.6 class models are really good at executing an existing plan down to a T. So quality becomes a function of how much you invest in the plan.
https://github.com/backnotprop/plannotator
It integrates with the CLI through hooks. completely local.
What’s wrong with you, people? Are you stupid?
Or, in CLAUDE.md have an instruction to follow AGENT.md - but this approach is quite unreliable.
These are solutions to a problem that shouldn’t exist in the first place. How else can one explain Anthropic’s reluctance to adhere to a widely adopted standard, if not as an attempt to build a walled garden around an otherwise great product?
They could potentially dumb it down further, but if they did that, it would hurt other use cases and competitors much more.
With stupidity like this what do they expect? It’s only a matter of time before people jump ship entirely.
After the Anthropic PMs have to delete their hundredth ticket about this issue, they will feel the need to fix it ... if only to stop the ticket deluge!
Map it to a workplace:
- Hey Joe, why did you stop adding code diff to your review requests?
- Most reviewers find it simpler. You can always run tcpdump on our shared drive to see what exactly was changed.
- I'm the only one reviewing your code in this company...
They could change course, obviously. But how does the saying go again -- it's easier for a camel to go through the eye of a needle, than for a VC funded tech startup to not enshittify.
The fact that LLM miss to read files is crucial for solving tasks. It does not matter that LLM later say "Yeah, I've fully read the specification and here is your code" if you check the log and it says: "Reading SPEC.md lines 1-400" <end_of_read>.
Overall, the complete log of interaction with the system should always be available, otherwise it is effectively a malware. That's not an exaggeration: consider that at any point of time any side part can spit out a prompt injection. Consider the use case: previously in xz-utils it was needed to sabotage the landlock kernel level sandbox, AND to exist in the memory of sshd, AND to be able to hijacking the RSA_public_decrypt. Now the only thing is needed - printf.
Boris's response here is the right move though. Acknowledging the miss and committing to a fix in the next release is how you build trust with a dev audience.
This is spreading like a plague: browser address bars are being trimmed down to nothing. Good luck figuring out which protocol you're using, or soon which website you are talking to. The TLS/SSL padlock is gone, so is the way to look into the site certificate (good luck doing that on recent Safari versions). Because users might be confused.
Well the users are not as dumb as you condescendingly make them out to be.
And if you really want to hide information, make it a config setting. Ask users if they want "dumbo mode" and see if they really do.
Also, was the padlock really such a problem? Did it really have to be removed? If not, perhaps another easily accessible way to access this data could be invented. Like, I don't know, a menu item perhaps?
I’ve been persistently dealing with the agent running in circles on itself when trying to fix bugs, not following directions fully and choosing to only accomplish partial requests, failing to compact and halting a session, and ignoring its MCP tooling and doing stupid things like writing cruddy python and osascripts unnecessarily.
I’ve been really curious about codex recently, but I’m so deep into Claude Code with multiple skills, agents, MCPs, and a skill router though.
Can anyone recommend an easy migration path to codex as a first time codex user from Claude code?
I care A LOT about the details, and I couldn't care less that they're cleaning up terminal output like this.
Seems like a dashboard mode toggle to run in a dedicated terminal would be a good candidate to move some of this complexity Anthropic seems to think “most” users can’t handle. When your product is increasing cognitive load the answer isn’t always to remove the complexity entirely. That decision in this case was clearly the wrong one.
I had used a Visa card to buy monthly Pro subscription. One day I ran out of credits so I go to buy extra credit. But my card is declined. I recheck my card limit and try again. Still declined.
To test the card I try extending the Pro subscription. It works. That's when I notice that my card has a security feature called "Secure by Visa". To complete transaction I need to submit OTP on a Visa page. I am redirected to this page while buying Pro subscription but not when trying to buy extra usage.
I open a ticket and mention all the details to Claude support. Even though I give them the full run down of the issue, they say "We have no way of knowing why your card was declined. You have to check with your bank".
Later I get hold of a Mastercard with similar OTP protection. It is called Mastercard Securecode. The OTP triggers on both subscription and extra usage page.
I share this finding with support as well. But the response is same - "We checked with our engineering team and we have no way of knowing why the other Visa card was declined. You have to check with your bank".
I just gave up trying to buy extra usage. So, I am not really surprised if they keep making the product worse.
Both Anthropic and OpenAI have been maintaining a high pace of releasing often poorly thought through new products and experimenting with features. A lot of their product releases show all the hallmarks of vibe coding: randomly breaking features, poor QA and testing on releases, etc.
OpenAI seems to have the upper hand in UX currently. Their products feel a bit more polished and they've clearly tried to up their game. Taking over Jony Ive's company a few months ago is a clear signal that they want to do better. The Codex AI desktop app was a clear step up from their web app and cli. I've been using both before that was released.
Both companies are spread very thin trying to do both end user and developer oriented products and features while keeping existing paying users happy as well. Both companies also have had a string of rushed product releases that kind of fizzled out: OpenAI's Atlas, which was a response to Anthropic's Comet. Neither of which seem to be very popular at this point. Several false starts with apps (OpenAI), Claude Cowork, etc. There are a lot of half formed product ideas there that than don't get the attention they deserve.
And it's not like MS, Google, and Apple are any better. If anything they are more hesitant and out of their depth here. They are all dancing around the hard issues here which are UX and security/trust models. Also, while coders get a lot of toys, nailing agentic tools for business users is proving to be a lot harder. Blanket access to everything via an agentic browser is not a viable solution. I can agentically code a structured document via latex or markdown. But the same tools are relatively useless in spreadsheets, presentations, and documents. And while you can do a lot of potentially interesting things if you surrender your inbox, the security failure modes around that remain a show stopping obstacle for wide adoption.
There's a lot of stage fright, hesitation, and immature product management in this sector. There's a bit of gold rush in terms of rapid experimentation. But as the stakes get higher, a lot of these companies are increasingly lacking the freedom to move as fast as needed. Fear of liability issues is preventing them to do a lot. Which is why most progress is concentrated around developer tools.
Meanwhile OpenCode is right there. (despite Anthropic efforts, you can still use it with a subscription) And you can tweak it any way you want...
That said, Cursor Composer is a lot faster and really nice for some tasks that don't require lots of context.
The value isn't just the models. Claude Code is notably better than (for example) OpenCode, even when using the same models. The plug-in system is also excellent, allowing me to build things like https://charleswiltgen.github.io/Axiom/ that everyone can benefit from.
Or, it decided it needs to get API documentation out and spends tens of thousands of tokens fetching every file in a repo with separate tool use instead of reading the documentation.
Profitable, if you are charging for token usage, I suspect.
But I’m reaching the point where I can’t recommend claude to people who are interesting in skeptically trying it out, because of the default model.
If they hide how the tool is accessing files (aka using tokens) and then charging us per token - how are we able to track loosely what our spend is?
I’m all for simplification of the UX. But when it’s helping to hide the main spend it feels shitty.
i bet in a week itll eat the whole 5hour throttle in one call too:P
ChatGPT or Gemini: I ask it what I wish to do, and show it the relevant code. It gives me a often-correct answer, and I paste it into my program.
Claude: I do the same, and it spends a lot of time thinking. When I check the window for the result, it's stalled with a question... asking to access a project or file that has nothing to do with the problem, and I didn't ask it to look for. Repeat several times until it solves the problem, or I give up with the questions.
i guess they were blocking OpenCode for a reason
this will put people to the test that use mainly Anthropic, to have a second look at the results from other models
Codex/Claude would like you to ignore both the code AND the process of creating the code.
You can control this behavior, so it's not a dealbreaker. But it shows a sort of optimism that skills make everything better. My experience is that skills are only useful for specific workflows, not as a way to broadly or generally enhance the LLM.
> “Searched for 1 pattern.”
Hit Ctrl-o like it mentions right there, and Claude Code will show you. Or RTFM and adjust Output Styles[1]. If you don't like these things, you can change them.
Like it or not, agentic coding is going mainstream and so they are going to tailor the default settings toward that wider mainstream audience.
It doesn't say "Read 3 files." though - it says "Read 3 files (ctrl+o to expand)" and you press ctrl+o and it expands the output to give you the detail.
It's a really useful feature to increase the signal to noise ratio where it's usually safe to do so.
I suspect the author simply needs to enable verbose mode output.
We carefully considered this change and feel it brings the most value to our users, and we hope you'll love chisel as much as we do.
One day, you guys are gonna learn not to tie your livelihoods to the whim of a corporation, but today isn't that day.
It is clever. its its best and worst feature.
Both browsers show “not secure” pretty prominently for non-TLS sites, and very loudly complain about sites with untrusted certificates, so the absence of either of those things signals a trusted cert, which is now the most common case by a very wide margin for me.
Usually I hate programming but it feels like a nice little tool to create
Nix makes it easy to package up esotheric patches reliably and reproducibly, claude lowers the cost of creating such patches, the only roadblocks Inforesee are legal.
For those of you who are still suckered in paying for it, why do you think the company would care how they abuse the existing users? You all took it the last time.
I am interested in the more abstract and general concept of: "People excessively feel that things are worse, even if they are not." And I see this A LOT in the AI/LLM area.
For instance, the claim that Claude Code, on the UX/DX side, is dumbed down seems to me absolutely not a reasonable take. The "hiding" of the file name being read is no longer being shown neither supports that claim, AND has to be seen in the context of Claude Code as a whole.
On the first point: Could one not make the argument that "not showing files read", is part of a more advanced abstraction layer, switching emphasis to something else in the UX experience? That could, by some, be seen as the overall package becoming more advanced and making choices as to what is presented for cognitive load. Secondly... it's not removed. It's just not default shown in non-verbose mode. As I understand it, you can just hit CTRL+O to see it again.
Secondly, even if it was done ONLY to be less for "power user focus," and more for dumb people (got to love the humility in the developer world), it's blindly obvious that you can't just mention ONE change as proof that Claude Code is dumbed down. And to me, it just does not compute to say that Claude Code feels dumbed down over the last patches. The amount of more advanced features, like seeing background tasks, the "option" selection feature, lifecycle hooks, sub-agents, agent swarms, skills—all of these have been released in just the last few months. I have used Claude Code since the very beginning, and it is just insane to claim that it's getting dumber as a tool. And this is just in relationship to the actual functionality, UX, and DX, not the LLM quality. But people see "I now have to hit CTRL+O to see files being read = DUMBED DOWN ENSHITFICATION!!!" I don't get it.
My point was simply... I'm much more interested in the psychological aspects driving everybody to predictably always claim that "things are getting worse," when it seems to not be the case. Be that in the exaggerated (but sometimes true) claims of model degradation, or as in this example of Claude Code getting dumbed down. What is driving this bias towards seeing and claiming things are getting worse, out of proportion to reality?
Or even shorter: why are we obsessed with the narrative of decline?
No affiliation, just a fan.
If you are talking about a CLI editor, then micro has hit the nail on quality UX
What if you hit ctrl+o?
I mean I get it I guess but I'm not nearly so passionate as anyone saying things about this
But personally I really love these new copy paste conventions, its the ctrl q convention which troubled me in ghostty but what I did was "ctrl > " write quit enter
https://github.com/micro-editor/micro/blob/master/runtime/he...
I may not be up to date with the latest & greatest on how to code with AI, but I noticed that as opposed to my more human in the loop style,
But FYI the blog post is not about the actual model being dumbed down, but the command line interface.
There's no conspiracy, though, other than more tokens consumed = more money, and they want that.
> What majority? The change just shipped and the only response it got is people complaining.
I'll refer you to the old image of the airplane with red dots on it. The people who don't have a problem with it are not complaining.
> People explained, repeatedly, that they wanted one specific thing: file paths and search patterns inline. Not a firehose of debug output.
Same as above. The reality is there are lots of people whose ideal case would be lots of different things, and you're seeking out the people who feel the same as you. I'm not saying you're wrong and these people don't exist, but you have to recognize that just because hundreds or thousands or tens of thousands of people want something from a product that is used by millions does not make it the right decision to give that thing to all of the users.
> Across multiple GitHub issues opened for this, all comments are pretty much saying the same thing: give us back the file paths, or at minimum, give us a toggle.
This is a thing that people love to suggest - I want a feature but you're telling me other people don't? Fine, just add a toggle! Problem solved!
This is not a good solution! Every single toggle you add creates more product complexity. More configurations you have to QA when you deploy a new feature. Larger codebase. There are cases for a toggle, but there is also a cost for adding one. It's very frequently the right call by the PM to decline the toggle, even if it seems like such an obvious solution to the user.
> The developer’s response to that?
> I want to hear folks’ feedback on what’s missing from verbose mode to make it the right approach for your use case.
> Read that again. Thirty people say “revert the change or give us a toggle.” The answer is “let me make verbose mode work for you instead.”
Come on - you have to realize that thirty people do not in any way comprise a meaningful sample of Claude Code users. The fact that thirty people want something is not a compelling case.
I'm a little miffed by this post because I've dealt with folks like this, who expect me as a PM to have empathy for what they want yet can't even begin to considering having empathy for me or the other users of the product.
> Fucking verbose mode.
Don't do this. Don't use profanity and talk to the person on the other side of this like they're an idiot because they're not doing what you want. It's childish.
You pay $20/month or maybe $100/month or maybe even $200/month. None of those amounts entitles you to demand features. You've made your suggestion and the people at Anthropic have clearly listened but made a different decision. You don't like it? You don't have to use the product.
You all are refining these models through their use, and the model owners will be the only ones with access to true models while you will be fed whatever degraded slop they give you.
You all are helping concentrate even more power in these sociopaths.
This is the true AI pilled version.
"docker sandbox run claude" in a recent version of docker is a super easy way to get started.
The alternatives are alienating users or dumbing down the software, both of which are worse for any serious professional product.
> The alternatives are alienating users or dumbing down the software, both of which are worse for any serious professional product.
I disagree that this is universally true. Alienating users is very frequently the right call. The alienated users never feel that way, but it's precisely the job of the PM to understand which users they want to build the product for and which ones they don't. You have to be fine alienating the latter group.
It may be that a dev implemented it, but it's the PM's job to make up excuses.
What's next? Calling us confused?
For example, applying diffs to files. Since the LLM uses tokenization for all its text input/output, sometimes the diffs it'll create to modify a file aren't quite right as it may slightly mess up the text which is before/after the change and/or might introduce a slight typo in text which is being removed, which may or may not cleanly apply in the edit. There's a variety of ways to deal with this but most of the agentic coding tools have this mostly solved now (I guess you could just copy their implementation?).
Also, sometimes the models will send you JSON or XML back from tool calls which isn't valid, so your tool will need to handle that.
These fun implementation details don't happen that often in a coding session, but they happen often enough that you'd probably get driven mad trying to use a tool which didn't handle them seamlessly if you're doing real work.
It's probably not enough to have answer-prompt -> tool call -> result critic -> apply or refine, there might be a specific thing they're doing when they fine tune the loop to the model, or they might even train the model to improve the existing loop.
You would have to first look at their agent loop and then code it up from scratch.
edit: There's a tool, i haven't used it in forever, i think it was netsaint(?) that let you sniff https in clear text with some kind of proxy. The enabling requirement is sniffing traffic on localhost iirc which would be the case with CC
You think a single person can do better? I don't think that's possible. Opencode is better than Claude Code and they also have perhaps even more manhours.
It's a collaboration thing, ever improving.
It's an experienced reality indeed, but THEN you create a narrative based on that. Obviously.
Experienced reality is, by definition, subjective and affected by filters for what you can, and how, experience things.
For instance, you can actually and truly experience something as bad, and then create a narrative around that. And you can be right, or you can be wrong in the narrative. Some narcissists experience themselves as a victim and unfairly treated, but everybody around them thinks the victim narrative is wrong, because they can clearly see that they are primarily at fault for their own situation.
So you just shifted the question to: "Why do people have a bias towards experience something as worsening, regardless of objective measures of quality"?
Sometimes people are right about something and sometimes the are not?
That's all I got out of it.
Start small, hit issues, fix them, add features, iterate, just like any other software.
There's also a handful of smaller open source agentic tools out there which you can start from, or just join their community, rather than writing your own.
ML related stuff isnt going to matter a ton since for most cases an LLM inference is you making an API call
web scraping is probably the most similar thing
Let's say we had objective data on things people say that we know are wrong regarding LLMs. The amount of people who WRONGLY say "It's getting worse" dwarfs the amount of people who WRONGLY think it has gotten better.
All I said is that I'm starting to get more interested in the psychological factors for this observation, the negativity bias, than actually investigating if the latest in a series of "OMG MODEL DEGRADATION" or "UI SUCKS NOW" posts; is actually true.
One of the hard things about building a product on an LLM is that the model frequently changes underneath you. Since we introduced Claude Code almost a year ago, Claude has gotten more intelligent, it runs for longer periods of time, and it is able to more agentically use more tools. This is one of the magical things about building on models, and also one of the things that makes it very hard. There's always a feeling that the model is outpacing what any given product is able to offer (ie. product overhang). We try very hard to keep up, and to deliver a UX that lets people experience the model in a way that is raw and low level, and maximally useful at the same time.
In particular, as agent trajectories get longer, the average conversation has more and more tool calls. When we released Claude Code, Sonnet 3.5 was able to run unattended for less than 30 seconds at a time before going off the rails; now, Opus 4.6 1-shots much of my code, often running for minutes, hours, and days at a time.
The amount of output this generates can quickly become overwhelming in a terminal, and is something we hear often from users. Terminals give us relatively few pixels to play with; they have a single font size; colors are not uniformly supported; in some terminal emulators, rendering is extremely slow. We want to make sure every user has a good experience, no matter what terminal they are using. This is important to us, because we want Claude Code to work everywhere, on any terminal, any OS, any environment.
Users give the model a prompt, and don't want to drown in a sea of log output in order to pick out what matters: specific tool calls, file edits, and so on, depending on the use case. From a design POV, this is a balance: we want to show you the most relevant information, while giving you a way to see more details when useful (ie. progressive disclosure). Over time, as the model continues to get more capable -- so trajectories become more correct on average -- and as conversations become even longer, we need to manage the amount of information we present in the default view to keep it from feeling overwhelming.
When we started Claude Code, it was just a few of us using it. Now, a large number of engineers rely on Claude Code to get their work done every day. We can no longer design for ourselves, and we rely heavily on community feedback to co-design the right experience. We cannot build the right things without that feedback. Yoshi rightly called out that often this iteration happens in the open. In this case in particular, we approached it intentionally, and dogfooded it internally for over a month to get the UX just right before releasing it; this resulted in an experience that most users preferred.
But we missed the mark for a subset of our users. To improve it, I went back and forth in the issue to understand what issues people were hitting with the new design, and shipped multiple rounds of changes to arrive at a good UX. We've built in the open in this way before, eg. when we iterated on the spinner UX, the todos tool UX, and for many other areas. We always want to hear from users so that we can make the product better.
The specific remaining issue Yoshi called out is reasonable. PR incoming in the next release to improve subagent output (I should have responded to the issue earlier, that's my miss).
Yoshi and others -- please keep the feedback coming. We want to hear it, and we genuinely want to improve the product in a way that gives great defaults for the majority of users, while being extremely hackable and customizable for everyone else.
To try it: /config > verbose, or --verbose.
Please keep the feedback coming. If there is anything else we can do to adjust verbose mode to do what you want, I'd love to hear.
shhh don't say that, they will never fix it if means you use less tokens.
I understand that I’m probably not the target audience if I want to actually step in and correct course, but it’s annoying.
Sighted users lost convenience. I lost the ability to trust the tool. There is no "glancing" at terminal output with a screen reader. There is no "progressive disclosure." The text is either spoken to me or it doesn't exist.
When you collapse file paths into "Read 3 files," I have no way to know what the agent is doing with my codebase without switching to verbose mode, which then dumps subagent transcripts, thinking traces, and full file contents into my audio stream. A sighted user can visually skip past that. I listen to every line sequentially.
You've created a situation where my options are "no information" or "all information." The middle ground that existed before, inline file paths and search patterns, was the accessible one.
This is not a power user preference. This is a basic accessibility regression. The fix is what everyone in this thread has been asking for: a BASIC BLOODY config flag to show file paths and search patterns inline. Not verbose mode surgery. A boolean.
Please just add the option.
And yes, I rewrote this with Claude to tone my anger and frustration down about 15 clicks from how I actually feel.
> Yoshi and others -- please keep the feedback coming. We want to hear it, and we genuinely want to improve the product in a way that gives great defaults for the majority of users, while being extremely hackable and customizable for everyone else.
I think an issue with 2550 upvotes, more than 4 times of the second-highest, is very clear feedback about your defaults and/or making it customizable.
Are you actually wondering, or just hoping to hear a confirmation of what you already know? Because the reason behind it is pretty clear, it doubles as both vendor lock-in and advertisement.
> The amount of output this generates can quickly become overwhelming in a terminal
If I use Opus 4.6, arguably the most verbose, over thinking model you've released to date, OpenCode handles it just the same as it does Sonnet 4.0.
OpenCode even allows me to toggle into subagent and task agents with their own output terminals that, if I am curious what is going on, I can very clearly see it.
All Claude-Code has done has turned the output into a black box so that I am forced to wait for it to finish to look at the final git diff. By then it's spent $5-10 working on a task, and threw away a lot of the context it took to get there. It showed "thinking" blocks that weren't particularly actionable, because it was mostly talking to itself that it can't do something because it goes against a rule, but it really wants to.
I'm actually frustrated with Code blazing through to the end without me able to see the transcript of the changes.
Funnily enough, both independently sided with the users, not the authors.
The core problem: --verbose was repurposed instead of adding a new toggle. Users who relied on verbose for debugging (thinking, hooks, subagent output) now have broken workflows - to fix a UX decision that shouldn't have shipped as default in the first place.
What should have been done:
/config
Show file paths: [on/off]
Verbose mode: [on/off] (unchanged)
Opus 4.6's parting thought: if you're building a developer tool powered by an AI that can reason about software design, maybe run your UX changes past it before shipping.
To be fair, your response explains the design philosophy well - longer trajectories, progressive disclosure, terminal constraints. All valid. But it still doesn't address the core point: why repurpose --verbose instead of adding a separate toggle? You can agree with the goal and still say the execution broke existing workflows.
But this one isn't? I'd call myself a professional. I use with tons of files across a wide range of projects and types of work.
To me file paths were an important aspect of understanding context of the work and of the context CC was gaining.
Now? It feels like running on a foggy street, never sure when the corner will come and I'll hit a fence or house.
Why not introduce a toggle? I'd happily add that to my alisases.
Edit: I forgot. I don't need better subagent output. Or even less output whrn watching thinking traces. I am happy to have full verbosity. There are cases where it's an important aspect.
More details here: https://news.ycombinator.com/item?id=46982177
Did you ever think that this may be Anthropic's goal? It is a waste for sure but it increases their revenue. Later on the old feature you were used to may resurface at a different tier so you'd have to pay up to get it.
https://martin.ankerl.com/2007/09/01/comprehensive-linux-ter...
Could the React rendering stack be optimised instead?
I just find that very hard to believe. Does anyone actually do anything with the output now? Or are they just crossing their fingers and hoping for the best?
If you are serious about this, I think there are so many ways you could clean up, simplify, and calm the Claude Code terminal experience already.
I am not a CC user, but an enthusiastic CC user generously spent an hour or two last week or so showing me how it worked and walking through an non-publicly-implemented Gwern.net frontend feature (some CSS/JS styling of poetry for mobile devices).
It was highly educational and interesting, and Claude got most of the way to something usable.
Yet I was shocked and appalled by the CC UI/UX itself: it felt like the fetal alcohol syndrome lovechild of a Las Vegas slot machine and Tiktok. I did not realize that all those jokes about how using CC was like 'crack' or 'ADHD' or 'gambling' were so on point, I thought they were more, well, metaphorical about the process as a whole. I have not used such a gross and distracting UI in... a long time. Everything was dancing and bouncing around and distracting me while telling me nothing. I wasted time staring at the update monitor trying to understand if "Prognosticating..." was different from "Fleeblegurbigating..." from "Reticulating splines...", while the asterisk bounces up and down, or the colored text fades in and out, all simultaneously, and most of the screen was wasted, and the whole thing took pains to put in as much fancy TUI nonsense as it could. An absolute waste, not whimsy, of pixels. (And I was a little concerned how much time we spent zoned out waiting on the whole shabang. I could feel the productivity leaving my body, minute by minute. How could I possibly focus on anything else while my little friendly bouncing asterisk might finish at any instant...?!) Some description of what files are being accessed seems like you could spare the pixels for them.
So I was impressed enough with the functionality to move it up my list, but also much of it made me think I should look into GPT Codex instead. It sounds like the interfaces there respect my time and attention more, rather than treating me like a Zoomer.
Please revert this
That's why I use your excellent VS Code extension. I have lots of screen space and it's trivial to scroll back there, if needed.
I would really like even more love given to this. When working with long-lived code bases it's important to understand what is happening. Lots of promising UX opportunities here. I see hints of this, but it seems like 80% is TBD.
Ideally you would open source the extension to really use the creativity of your developer user base. ;)
It might be worth considering a "verbose level" type setting with a selection of levels that describe the level of verbosity. Effectively, use a select menu instead of a boolean when one boolean state is actually multiple nested states.
Edit: I realised my use of "verbose" and "verbosity" here is it self ironically verbose, sorry!
As others have said - 'reading 10 files' is useless information - we want to be able to see at a glance where it is and what it's doing, so that we can re-direct if necessary.
With the release of Cowork, couldn't Claude Code double down on needs of engineers?
Ooo... ooo! I know what this is a reference to!
If that's the case, it's important to asses wether it'll be consistent when operating on a higher level, less dependent on the software layer that governs the agent. Otherwise it'll risk Claude also becoming more erratic.
Of course all the logs can’t be streamed to a terminal. Why would they need to be? Every logging system out there allows multiple stream handlers with different configurations.
Do whatever reasonable defaults you think make sense for the TUI (with some basic configuration). But then I should also be able to give Claude-code a file descriptor and a different set of config optios, and you can stream all the logs there. Then I can vibe-code whatever view filter I want on top of that, or heck, have a SLM sub-agent filter it all for me.
I could do this myself with some proxy / packet capture nonsense, but then you’d just move fast and break my things again.
I’m also constantly frustrated by the fancier models making wrong assumptions in brownfield projects and creating a big mess instead of asking me follow-up questions. Opus is like the world’s shittiest intern… I think a lot of that is upstream of you, but certainly not all of it. There could be a config option to vary the system prompt to encourage more elicitation.
I love the product you’ve built, so all due respect there, but I also know the stench of enshittification when I smell it. You’re programmers, you know how logging is supposed to work. You know MCP has provided a lot of these basic primitives and they’re deliberately absent from claude code. We’ve all seen a product get ratfucked internally by a product manager who copied the playbook of how Prabhakar Raghavan ruined google search.
The open source community is behind at the moment, but they’ll catch up fast. Open always beats closed in the long run. Just look at OpenAI’s fall into disgrace.
This is verifiable bullshit. Unless you explicitly explain how it "runs for days" since Opus's context window is incapable of handling even relatively large CLAUDE.md files.
> The amount of output this generates can quickly become overwhelming in a terminal, and is something we hear often from users. Terminals give us relatively few pixels to play with; they have a single font size; colors are not uniformly supported; in some terminal emulators, rendering is extremely slow.
No. It's your incapability as an engineer that limits this. And you and your engineers getting high on your own supply. Hence you need 16ms to draw a couple of characters on screen and call it a tiny game engine [1] For which your team was rightfully ridiculed.
> But we missed the mark for a subset of our users. To improve it,
AI-written corporate nothingspeak.
Maybe "AI IDEs" will gain ground in the future, e.g. vibe-kanban
Unfortunately, vibe coders cannot do that anymore.
It doesn't compose with any other command line program and the terminal interface is limiting.
I'm surprised nobody has yet made a coding assistant that runs in the browser or as a standalone app. At this point it doesn't really need to integrate with my text editor or IDE.
Product manager here. Cynically, this is classic product management: simplify and remove useful information under the guise of 'improving the user experience' or perhaps minimalism if you're more overt about your influences.
It's something that as an industry we should be over by now.
It requires deep understanding of customer usage in order not to make this mistake. It is _really easy_ to think you are making improvements by hiding information if you do not understand why that information is perceived as valuable. Many people have been taught that streamlining and removal is positive. It's even easier if you have non-expert users getting attention. All of us here at HN will have seen UIs where this has occurred.
https://github.com/anthropics/claude-code/issues/15263
https://github.com/anthropics/claude-code/issues/9099
https://github.com/anthropics/claude-code/issues/8371
It's very clear that Anthropic doesn't really want to expose the secret sauce to end users. I have to patch Claude every release to bring this functionality back.
And then this. They want to own your dev workflow and for some reason believe Claude code is special enough to be closed source. The react TUI is kinda a nightmare to deal with I bet.
I will say, very happy with the improvements made to Codex 5.3. I’ve been spending A LOT more time with codex and the entire agent toolchain is OSS.
Not sure what anthropic’s plan is, but I haven’t been a fan of their moves in the past month and a half.
for example Amp "feels" much better. Also like in Amp how I can just send the message whenever and it doesn't get queued
* I know, lots of "feels" in there..
https://www.nytimes.com/2025/01/08/technology/sam-altman-sis...
DEVELOPERS, DEVELOPERS, DEVELOPERS, DEVELOPERS
I write mainly out of the hope that some Anthropic employees read this: you need an internal crusade to fight these impulses. Take the high road in the short-term and you may avoid being disrupted in the long-term. It's a culture issue.
Probably your strongest tool is specifically educating people about the history. Microsoft in the late 90s and early 00s was completely dominant, but from today's perspective it's very clear: they made some fundamental choices that didn't age well. As a result, DX on Windows is still not great, even if Visual Studio has the best features, and people with taste by and large prefer Linux.
Apple made an extremely strategic choice: rebuild the OS around BSD, which set them up to align with Linux (the language of servers). The question is: why? Go find out.
The difference is a matter of sensibility, and a matter of allowing that sensibility to exist and flourish in the business.
I understand the article writers frustration. He liked a thing about a product he uses and they changed the product. He is feeling angry and he is expressing that anger and others are sharing in that.
And I'm part of another group of people. I would notice the files being searched without too much interest. Since I pay a monthly rate, I don't care about optimizing tokens. I only care about the quality of the final output.
I think the larger issue is that programmers are feeling like we are losing control. At first we're like, I'll let it auto-complete but no more. Then it was, I'll let it scaffold a project but not more. Each step we are ceding ground. It is strange to watch someone finally break on "They removed the names of the files the agent was operating on". Of all of the lost points of control this one seems so trivial. But every camels back has a breaking point and we can't judge the straw that does it.
Maybe Claude Code web or desktop could be targeted to these new vibe coders instead? These folks often don't know how simple bash commands work so the terminal is the wrong UX anyway. Bash as a tool is just very powerful for any agentic experience.
It still does what I need so I'm okay with it, but I'm also on the $20 plan so it's not that big of a worry for me.
I did sense that the big wave of companies is hitting Anthropic's wallet. If you hadn't realized, a LOT of companies switched to Claude. No idea why, and this is coming from someone who loves Claude Code.
Anyway, getting some transparency on this would be nice.
It is entirely due to Opus 4.5 being an inflection point codingwise over previous LLMs. Most of the buzz there has been organic word of mouth due to how strong it is.
Opus 4.5 is expensive to put it mildly, which makes Claude Code more compelling. But even now, token providers like Openrouter have Opus 4.5 as one of its most popular models despite the price.
Why would you even watch a YouTube video with ads?
There are ad blockers, sponsor segment blockers, etc. If you use them, it will block almost every kind of YouTube ad.
Use the pi coding agent. Bare-bones context, easy to hack.
Ads in ChatGPT. Removing features from Claude Code. I think we're just beginning to face the music. It's also funny that how Google "invented" ad injection in replies with real-time auction capabilities, yet OpenAI would be the first implementer of it. It's similar to how transformers played out.
For me, that's another "popcorn time". I don't use any of these to any capacity, except Gemini, which I seldom use to ask stuff when deep diving in web doesn't give any meaningful results. The last question I asked managed to return only one (but interestingly correct) reference, which I followed and continued my research from there.
Regarding the thoughts: it also allows me to detect problematic paths it takes, like when it can't find a file.
For example today I was working on a project that depends on another project, managed by another agent. While refactoring my code it noticed that it needs to see what this command is which it is invoking, so it even went so far as to search through vs code's user data to find the recent files history if it can find out more about that command... I stopped it and told it that if it has problems, it should tell me. It explained it can't find that file, i gave it the paths and tokens were saved. Note that in that session I was manually approving all commands, but then rejected the one in the data dir.
Why dumb it down?
TIL that there's an especially apt xkcd comic for this scenario: "Zealous Autoconfig"
Set minimal defaults to keep output clean, but let users pick and choose items to output across several levels of verbosity, similar to tcpdump, Ansible, etc. (-v to -vvvvv).
I know businesses are obsessed with providing Apple-like "experiences", where the product is so refined there's just "the one way" to magically do things, but that's not going to work for a coding agent. It needs to be a unix-like experience, where the app can be customized to fit your bespoke workflow, and opening the man page does critical damage unless you're a wizard.
LLMs are already a magic box, which upsets many people. It'll be a shame if Anthropic alienates their core fan base of SWEs by making things more magical.
There are no vibes in “I am looking at files and searching for things” so I have zero weight to assign to your decision quality up until the point where it tells me the evals passed at 100%.
Your agent is not good enough. I trust it like I trust a toddler not to fall into a swimming pool. It’s not trying to, but enough time around the pool and it is going to happen, so I am watching the whole time, and I might even let it fall in if I think it can get itself out.
When you're building agents that interact with real environments (browsers, codebases, APIs), the single hardest thing to get right isn't the model's reasoning. It's giving the operator enough visibility into what the agent is actually doing without drowning them in noise. There's a narrow band between "Read 3 files" (useless) and a full thinking trace dump (unusable), and finding it requires treating observability as a first-class design problem, not a verbosity slider.
The frustrating part is that Anthropic clearly understands this in other contexts. Their own research on agent safety talks extensively about the need for human oversight of autonomous actions. But the moment it's their own product, the instinct is to simplify away the exact information that makes oversight possible.
The people pinning to 2.1.19 aren't being difficult. They're telling you that when an agent touches my codebase, I need to know which files it read and what it searched for — not because I want to micromanage, but because that's literally the minimum viable audit trail. Take that away and you're asking users to trust a black box that edits production code.
The other fact pattern is their CLI is not open source, so we can't go in and change it ourselves. We shouldn't have to. They have also locked down OpenCode and while there are hacks available, I shouldn't have to resort to such cat and mouse games as someone who pays $200/month for a premium service.
I'm aggressively exploring other options, and it's only a matter of if -- not when, one surfaces.
I mean I hope it's just a single developer being stubborn rather than guidance from management asking everyone to simplify Claude Code for maximum mass appeal. But I agree otherwise, it's telling.
> Compacting fails when the thread is very large
> We fixed it.
> No you did not
> Yes now it auto compacts all messages.
> Ok but we don't want compaction when the thread isn't large, plus, it still fails when the compacted thread is too large
> ...
> Compacting fails when the thread is very large
Flips coin, it is Heads
> We fixed it.
> No you did not
Flips coin, it is Tails
> Yes now it auto compacts all messages.
Flips coin, it is Heads
> Ok but we don't want compaction when the thread isn't large, plus, it still fails when the compacted thread is too large
Flips coin, it is Grapefruit
> ...
Congratulations on a vibe solution, if you are unhappy with the frequency of isomorphic plagiarism... the vendor still has your money and new data =3
Often a codebase ends up with non-authoritative references for things (e.g. docs out of sync with implementation, prototype vs "real" version), and the proper solution is to fix and/or document that divergence. But let's face it, that doesn't always happen. When the AI reads from the wrong source it only makes things worse, and when you can't see what it's reading it's harder to even notice that it's going off track.
Ah, the old "you're holding it wrong."
There is almost no value in watching the stream of intermediate tokens. There's no need to micromanage the agent's steps. Just monitor the artifact and insist the LLM summarizes findings in plain English.
If it can't explain the proposed change coherently, it can't code it coherently either. `git restore .`
I find it much more effective to throw away bad sessions, try a new prompt than to massage the existing context swamp.
True vibe coders don't care about this.
I like that people who were afraid of CLIs perhaps are now warming up to them through tools like Claude Code but I don't think it means the interfaces should be simplified and dumbed down for them as the primary audience.
Sure you can press CTRL+O, but that's not realtime and you have to toggle between that and your current real time activity. Plus it's often laggy as hell.
I'm using it for converting all of the userspace bcachefs code to Rust right now, and it's going incredibly smoothly. The trick is just to think of it like a junior engineer - a smart, fast junior engineer, but lacking in experience and big picture thinking.
But if you were vibe coding and YOLOing before Claude, all those bad habits are catching up with you suuuuuuuuuuuper hard right now :)
It's a huge shift, but we need to start thinking of AI-tools as developer tools, just like a formatter, linter, or IDE would be.
The right move is diversity. Just like diversity of editors/IDEs. We need good open source claude code alternatives.
As a SE with over 15 years' professional experience, I find myself pointing out dumb mistakes to even the best frontier models in my coding agents, to refine the ouput. A "coder" who is not doing this on the regular is only a tool of their tool.
(in my mental model, a "vibe coder" does not do this, or at least does not do it regularly)
I do wonder if there is going to be much of a difference between using Claude Code vs. Copilot CLI when using the same models.
I’m also at MS, not (yet?) using Claude Code at work and pondering precisely the same question.
I tend to disagree. While I don't see meaningful _reasoning power_ between frontier models, I do see differences in the way they interact with my prompts.
I use exclusively Anthropic models because my interactions with GPT are annoying:
- Sonnet/Opus behave like a mix of a diligent intern, or a peer. It does the work, doesn't talk too much, gives answers, etc.
- GPT is overly chatty, it borderline calls me "bro", tend to brush issues I raise "it should be good enough for general use", etc.
- I find that GPT hardly ever steps back when diagnosing issues. It picks a possible cause, and enters a rabbit hole of increasingly hacky / spurious solutions. Opus/Sonnet is often to step back when the complexity increases too much, and dig an alternative.
- I find Opus/Sonnet to be "lazy" recently. Instead of systematically doing an accurate search before answering, it tries to "guess", and I have to spot it and directly tell it to "search for the precise specification and do not guess". Often it would tell me "you should do this and that", and I have to tell it "no, you do it". I wonder if it was done to reduce the number of web searches or compute that it uses unless the user explicitly asks.
I subscribe to max rn. Tons of money. Anthropic’s Super Bowl ads were shit, not letting us use open code was shit, and this is more shit. Might only be a single straw left before I go to codex (no one’s complaining about it. And the openclaw creator prefers it)
This dev is clearly writing his reply with Claude and sounding way too corpo. This feels like how school teachers would talk to you. Your response in its length was genuinely insulting. Everyone knows how to generate text with AI now and you’re doing a terrible job at it. You can even see the emdash attempt (markdown renders two normal dashes as an emdash).
This was his prompt “read this blog post, familiarize yourself with the mentioned GitHub issue and make a response on behalf of Anthropic.” He then added a little bit at the end when he realized the response didn’t answer the question and got so to fix the grammar and spelling on that.
Your response is appropriate for the masses. But we’re not. We’re the so called hackers and read right through the bs. It’s not even about the feature being gone anymore.
There is a principle we uphold as “hackers” that doesn’t align with this that pisses people off a lot more than you think. I can’t really put my finger on it maybe someone can help me out.
PS About the Super Bowl ads. Anyone that knows the story knows they’re exaggerated. (In the general public outside of Silicon Valley it’s like a 50/50 split or something about people liking or disliking AI as a whole rn. OpenAI is doing way more to help the case (not saying ads are a good thing). ) Open ai used to feel like the bad guy now it’s kinda shifting to anthropic. This, the ads and open code are all examples of it. (I especially recommend people watch the anthropic and open ai Super Bowl ads back to back)
> You can even see the emdash attempt (markdown renders two normal dashes as an emdash)
He says he wrote it all manually.[0] Obviously I can't know if that's true, but I do think your internal AI detector is at least overconfident. For example, some of us have been regularly using the double hyphen since long before the LLM era. (In Word, it auto-corrects to an en dash, or to an em dash if it's not surrounded by spaces. In plain text, it's the best looking easily-typable alternative to a dash. AFAICT, it's not actually used for dashes in CommonMark Markdown.)
The rest is more subjective, but there are some things Claude would be unlikely to write (like the parenthetical "(ie. progressive disclosure)" -- it would write "i.e." with both dots, and it would probably follow it with a comma). Of course those could all be intentional obfuscations or minimal human edits, but IMO you are conflating the corporate communications vibe with the LLM vibe.
Edit: I can't post anymore today apparently because of dang. If you post a comment about a bad terminal at least tell us about the rendering issues.
As someone who finds formal language a natural and better interface for controlling a computer, can you explain how and why you actually hate it? I mean not stuff like lack of discoverability, because you use a shell that lacks completion and documentation, that have been common for decades, I get those downsides, but why do you detest it in principle?
Can we please move the "Extended Thinking" icon back to the left side of claude desktop, near the research and web search icons? What used to be one click is now three.
use your own words!
i would rather read the prompt.
How can that be true, when you're deliberately and repeatedly telling devs (the community you claim to listen to) that you know better than they do? They're telling you exactly what they want, and you're telling them, "Nah." That isn't listening. You understand that, right?
And it shouldn’t need to be said, but the words that appear on the screen are from an actual person with, you know, feelings.
Arrogant and clueless, not exactly who I want to give my money to when I know what enshitification is.
They have horrible instincts and are completely clueless. You need to move them away from a public-facing role. It honestly looks so bad, it looks so bad that it suggests nepotism and internal dysfunction to have such a poor response.
This is not the kind of mistake someone makes innocently, it's a window into a worldview that's made me switch to gemini and reactivate cursor as a backup because it's only going to get worse from here.
The problem is not the initial change (which you would rapidly realize was a big deal to a huge number of your users) but how high-handed and incompetent the initial response was. Nobody's saying they should be fired, but they've failed in public in a huge way and should step back for a long time.
It should be a fad gone by at this point, but people never learn. Here's what to do instead: Find your most socially competent engineer, and have them talk to users a couple times a month. Just saved you thousands or millions in salaries, and you have a better chance of making things that your users actually want.
The problem is, it's hard to measure how good a PM is, even harder than for engineers. The instinct is to use product KPI's to do so, but especially at BigTech company, distribution advantages and traction of previous products will be the dominant factor here, and the best way of raising many product KPI's are actually user-hostile. Someone who has been a successful FAANG engineer who goes to a startup might lean towards over-engineering, but at least they should be sharp on the fundamentals. Someone who has been a successful FAANG PM might actually have no idea how to get PMF.
> Here's what to do instead: Find your most socially competent engineer, and have them talk to users a couple times a month
This is actually a great idea, but what will happen is this socially competent engineer will soon have a new full-time job gathering those insights, coalescing them into actionable product changes, persuading the rest of the org to adopt those changes, and making sure the original user insights make it into the product. Voila: you've re-invented product management.
But I actually think it's good to source PM's from people who've been engineers for a few years. PM's used to come from a technical background; Google famously gave entry-level coding tests to PM's well into the '10s. I dunno when it became more fashionable to hire MBA's and consultants into this role, but it may have been a mistake.
But not lately. Lately it’s been people who have very little relevant domain expertise, zero interest in putting in the time to develop said expertise beyond just cataloguing and regurgitating feedback from the customers they like most on a personal level, and seem to mostly have only been selected for the position because they are really good at office politics.
But I think it’s not entirely their fault. What I’ve also noticed is that, when I was on teams with really elective product managers, we also had a full time project manager. That possibly freed up a lot of the product manager’s time. One person to be good at the tactical so the other can be good at the strategic.
Since project managers have become passé, though, I think the product managers are just stretched too thin. Which sets up bad incentive structures: it’s impossible to actually do the job well anymore, so of course the only ones who survive are the office politicians who are really good at gladhanding the right people and shifting blame when things don’t go well.
In most of my engineering jobs, the Product Managers were much closer to our users than the engineers.
Good product managers are very valuable. There are a lot of bad ones carrying the product manager title because it was viewed as the easy way to get a job in tech without having to know how to program, but smart companies are getting better at filtering them out.
> Find your most socially competent engineer, and have them talk to users a couple times a month
Every single time I've seen this tried, it turns into a situation where one or two highly vocal customers capture the engineering team's direction and steer the product toward their personal needs. It's the same thing that happens when the sales people start feeding requests from their customers into the roadmap.
For example, I had one product manager who made themselves irrelevant because they wouldn't work with sales. The company needed to sell the product to pay us, and sales talked with potential buyers about what might swing their purchase decision and what they would pay extra for. Since the PM only talked to users and ignored sales when doing product design and product roadmaps, the way sales input got integrated into product development is that we frequently got top-down directives from management to prioritize one-off requests from sales over the roadmap. Needless to say, this didn't lead to a cohesive and easy-to-understand product.
Before I saw that PM failing, I hadn't thought about the relationship between product and sales.
> Find your most socially competent engineer,
These usually get promoted to product management anyway, so this isn't a new thought.
I agree completely that these are the important qualifications to be setting direction for a product.
> Find your most socially competent engineer, and have them talk to users a couple times a month.
This doesn't necessarily follow from the above, but in Anthropic's case specifically, where the users are software engineers, it probably would have worked better than whatever they have going on now.
In general, it's probably better to have domain experts doing product management, as opposed to someone who is trained in product management.
Unfortunately, he’s already two of our SEs and the CTO and we’re starting to run low on coders.
What are we going to do when we need a customer success manager or a profserv team?
Make the application configurable. Developers like to tinker with their tools.
I agree it's a mistake, but I don't believe that it's viewed that way by anyone making the decision to do it.
I think we can be more charitable. Don't you see, even here on HN, people constantly asking for software that is less bloated, that does fewer things but does them better, that code is cost, and every piece of complexity is something that needs to be maintained?
As features keep getting added, it is necessary to revisit where the UX is "too much" and so things need to be hidden, e.g. menu commands need to be grouped in a submenu, what was toolbar functionality now belongs in a dialog, reporting needs to be limited to a verbose mode, etc.
Obviously product teams get it wrong sometimes, users complain, and if enough users complain, then it's brought back, or a toggle to enable it.
There's nothing to be cynical about, and it's not something we "should be over by now." It's just humans doing their best to strike the balance between a UX that provides enough information to be useful without so much information that it overwhelms and distracts. Obviously any single instance isn't usually enough to overwhelm and distract, but in aggregate they do, so PM's and designers try to be vigilant to simplify wherever possible. But they're only human, sometimes they'll get it wrong (like maybe here), and then they fix it.
Here lies the problem
Your first users were here without all the added features. It's very likely they didn't need those features to use the software. Then you add new features and clutter it... then remove features from the UX that they were using.
Cynically, it's a vibe coded mess and the "programmers" at Anthropic can't figure out how to put it back.
More cynically, Anthropic management is trying to hide anything that people could map to token count (aka money) so that they can start jiggling the usage numbers to extract more money from us.
I was recently involved with a company that wanted us to develop a product that would be disruptive enough to enter an established market, make waves and shock it.
We did just that. We ran a deep survey of all competing products, bought a bunch of them, studied absolutely everything about them, how they were used and their users. Armed with that information, we produced a set of specifications and user experience requirements that far exceeded anything in the market.
We got green-lit to deliver a set of prototypes to present at a trade show. We did that.
The prototypes were presented and they truly blew everyone away. Blogs, vlogs, users, everyone absolutely loved what we created and the sense was that this was a winning product.
And then came reality. Neither the product manager nor the CTO (and we could add the CEO and CFO to the list) had enough understanding and experience in the domain to take the prototypes to market. It would easily have required a year or two of learning before they could function in that domain.
What did they do? They dumbed down the product specification to force it into what they understood and what engineering building blocks they already had. Square peg solidly and violently pounded into a round hole.
The outcome? Oh, they built a product alright. They sure did. And it flopped, horribly flopped, as soon as it was introduced and made available. Nobody wanted it. It was not competitive. It offered nothing disruptive. It was a bad clone of everything already occupying space in that ecosystem. Game over.
The point is: Technology companies are not immune to human failings, ego, protectionism/turf guarding, bad decisions, bad management, etc.
When someone says something like "I am not sure that's a good idea for a startup. There's competition." My first though is: Never assume that competitors know what they are doing, are capable and always make the right decisions without making mistakes. You don't always need a better product, you need better execution.
> The point is: Technology companies are not immune to human failings, ego, protectionism/turf guarding, bad decisions, bad management, etc.
They only accidentally succeed in spite of those things. They have those things more than existing businesses precisely because having too much money masks the pressures that would force solid execution and results. When you have 80% profit margins, you can show up drunk.
Over the past ten years or so the increasing de-featuring of software under the guise of 'simplification' has become a critical issue for power users. For any GUI apps which have a mixed base of consumer and power users, I mostly don't update them anymore because they're as likely to get net worse vs better.
It's weird that companies like MSFT seem puzzled why so many users refuse to update Windows or Office to major new feature versions.
We have by now taught them about good information density.
Like, the permission pages, if you look at them just once, kinda look like bad 90s UIs. They throw a crapton of information at you.
But they contain a lot of smart things you only realize when actually using it from an admin perspective. Easy comparison of group permissions by keeping sorting orders and colors stable, so you can toggle between groups and just visually match what's different, because colors change. Highlights of edge cases here and there. SSO information around there as well. Loads of frontloaded necessary info with optional information behind various places.
You can move seriously fast in that interface once you understand it.
Parts of the company hate it for not being user friendly. I just got a mail that a customer admin was able to setup SSO in 15 minutes and debug 2 mapping issues in another 10 and now they are production ready.
Not at all cynically, this is classic product management - simplify by removing information that is useful to some users but not others.
We shouldn't be over it by now. It's good to think carefully about how you're using space in your UI and what you're presenting to the user.
You're saying it's bad because they removed useful information, but then why isn't Anthropic's suggestion of using verbose mode a good solution? Presumably the answer is because in addition to containing useful information, it also clutters the UI with a bunch of information the user doesn't want.
Same thing's true here - there are people who want to see the level of detail that the author wants and others for whom it's not useful and just takes up space.
> It requires deep understanding of customer usage in order not to make this mistake.
It requires deep understanding of customer usage to know whether it's a mistake at all, though. Anthropic has a lot deeper understanding of the usage of Claude Code than you or I or the author. I can't say for sure that they're using that information well, but since you're a PM I have to imagine that there's been some time when you made a decision that some subset of users didn't like but was right for the product, because you had a better understanding of the full scope of usage by your entire userbase than they did. Why not at least entertain the idea that the same thing is true here?
The notifications act as an overall progress bar and give you a general sense of what Claude Code is doing: is it looking in the relevant part of your codebase, or has it gotten distracted by some unused, vendored-in code?
"Read 2 files" is fine as a progress indicator but is too vague for anything else. "Read foo.cpp and bar.h" takes almost the same amount of visual space, but fulfills both purposes. You might want to fold long lists of files (5? 15?) but that seems like the perfect place for a user-settable option.
Software developers like customizable tools.
That's why IDEs still have "vim keybindings" and many other options.
Your user is highly skilled - let him decide what he wants to see.
Because reading through hundreds of lines verbose output is not a solution to the problem of "I used to be able to see _at a glance_ what files were being touched and what search patterns were being used but now I can't".
I personally love that the model tells me what file it has read because I know whether or not it's headed in the generally right direction that I intended. Anthropic has no way of knowing I feel this way.
PM1> Looks like a PM who is out of touch with what the developers want. Easy mistake to make.
PM2> Anthropic knows better than this developer. The developer is probably wrong.
I don't know for sure what the best decision is here, I've barely used CC. Neither does PM1 nor PM2, but PM2 is being awfully dismissive of the opinion of a user in the target audience. PM1 is probably putting a bit too much weight on Developer's opinion, but I fully agree with "All of us... have seen UIs where this has occurred." Yes, we have. I personally greatly appreciate a PM who listens and responds quickly to negative feedback on changes like this, especially "streamlining" and "reducing clutter" type changes since they're so easy to get wrong (as PM1 says).
> It's good to think carefully about how you're using space in your UI and what you're presenting to the user.
I agree. It's also good to have the humility to know that your subjective opinion as someone not in the target audience even if you're designing the product is less informed in many ways than that of your users.
----
Personally, I get creeped out by how many things CC is doing and tokens it's burning in the background. It has a strong "trust me bro" vibe that I dislike. That's probably common to all agent systems; I haven't used enough to know.
People who toggle debug will get "full" access and those who dont care, probably won't notice if their LLM us is degraded.
It seems a pure market segmenting prior to a "shrinkflation" approach to cost management.
If Claude Code can replace an engineer, it should cost just a bit less than an engineer, not half as much.
That means the pricing is going to be competitive. You may still get your wish though, but instead of the price of an engineer remaining the same, it will cut itself down by 95%.
https://github.com/shepherdjerred/monorepo/tree/main/package...
Meanwhile, I am observing precisely how VS+Copilot works in my OAI logs with zero friction. Plug in your own API key and you can MITM everything via the provider's logging features.
I checked with ccusage (a cli tool that checks how much your Claude subscription tokens would have cost via the API).
My $200 a month subscription would have cost me more than $3000. The highest single day would have cost more than $300.
Gemini is cheaper, but not by much.
To other actors who want to train a distilled version of Claude, more likely.
1: https://blog.devgenius.io/you-might-be-breaking-claudes-tos-...
It's amazing how much other agentic tools suck in comparison to Claude Code. I'd love to have a proper alternative. But they all suck. I keep trying them every few months and keep running back to Claude Code.
Just yesterday I installed Cursor and Codex, and removed both after a few hours.
Cursor disrespected my setting to ask before editing files. Codex renamed my tabs after I had named them. It also went ahead and edited a bunch of my files after a fresh install without asking me. The heck, the default behavior should have been to seek permission at least the first time.
OpenCode does not allow me to scrollback and edit a prior prompt for reuse. It also keeps throwing up all kinds of weird errors, especially when I'm trying to use free or lower cost models.
Gemini CLI reads strange Python files when I'm working on a Node.js project, what the heck. It also never fixed the diff display issues in the terminal; It's always so difficult for me to actually see what edits it is actually trying to make before it makes it. It also frequently throws random internal errors.
At this point, I'm not sure we'll be seeing a proper competitor to Claude Code anytime soon.
GitHub Codespaces has a critical bug that makes the copilot terminal integration unusable after 1 prompt, but the company has no idea, because there is no clear way to report it from the product, no customer support funnel, etc. There's 10 upvotes on a poorly-written sorta-related GH issue and no company response. People are paying for this feature and it's just broken.
Anthropic is the market leader for advanced AI coding with no serious competitor currently very close and they are preparing to IPO this year. This year is a transition year. The period where every decision would default toward delighting users and increasing perceived value is ending. By next year they'll be fully on the quarterly Wall Street grind of min/maxing every decision to extract the highest possible profit from customers at the lowest possible cost.
This path is inevitable and unavoidable, even with the most well-intentioned management and employees.
Some greatest hits:
- CoreAudio, Mac OS memory management, kernel in general, and many other decisions
- Google's internal dev tooling, Go, and Chrome (at least, in its day)
- C#, .NET, and Typescript (even Microsoft does good work)
One of the hallmarks of heroic engineering work is that everyone takes it for granted afterward. Open source browsers that work, audio that just works, successors to C/C++ with actual support and adoption, operating systems that respond gracefully under load, etc. ... none of these things were guaranteed, or directly aligned with short-term financial incentives. Now, we just assume they're a requirement.
Part of the "sensibility" I'm talking about is seeking to build things that are so boring and reliable that nobody notices them anymore.
I'm guessing you're not aware of how their newest game, Starfield, was received. In the long term, that direction did not work out for them at all.
Those specific logs are essentially a prop anyways. Removing them makes it harder to LARP as an active participant; it forces the realization that "we" are now just passive observers.
If it sounds strange your theory might be wrong.
Telepsychology is one of the lowest forms of response.
Those are fightin’ words as someone who has dumped more hours than I can count into Skyrim but…
I had never heard of this game, but it has a lot going for it (source engine) and I watched a little of the gameplay you linked and I’m intrigued. I’m probably gonna pick this up for the steam deck.
A friend recommended the Might and Magic games to me a long time ago and I bought them off GoG, but wasn’t a fan of the gameplay and just couldn’t get hooked. This looks very different from what I remember (probably because this is a very different game from the earlier ones).
Thank you for mentioning this game!
On the other end are the hardcore user orchestrating a bunch of agents, not sitting there watching one run, so they don’t care about these logs at all
In the middle are the engineers sitting there watching the agent go
Or, it could serve as a textbook example how to make your real future long term customers (=fluent coders) angry… what a strategy :)
Yes, it was easier. But it dumbed down a generation of developers.
It took them two decades to try to come up with Powershell, but it was too late.
Meanwhile all evidence is that the true value of these tools is in their ability to augment & super-charge competent software engineers, not replace them.
Meanwhile the quality of Claude Code the tool itself is a bit of a damning indictment of their philosophy.
Give me a team of experienced sharp diligent engineers with these coding tools and we can make absolutely amazing things. But newbie product manager with no software engineering fundamentals issuing prompts will make a mess.
I can see it even in my own work -- when I venture into doing frontend eng using these tools the results look good but often have reliability issues. Because my background/specialization is in systems, embedded & backend work -- I'm not good at reviewing the React etc code it makes.
The whole company also has this meme about AI safety and some sort of fear-mongering about the models every few months. It's basically a smokescreen for normies and other midwits to make it look more mysterious and advanced than it really is. OOOOH IT'S GOING TO BREAK OUT! IT KNOWS IT'S BEING EVALUATED!
I bet there are some true believers in Anthropic too, people who think themselves too smart to believe in God so they replaced it with AI instead but all the same hopes are there, eg. Amodei preaching about AI doubling the human lifespan. In religion we usually talk about heaven.
Not a chance.
If anything, the reverse, in that it devalues engineering. For most, LLMs are a path to an end-product without the bother or effort of understanding. No different than paid engineers were, but even better because you don't have to talk to engineers or pay them.
The sparks of genuine curiosity here are a rounding error.
There is a reason why nowadays games start to help massively if the player gets stuck.
Programmers are just jealous that they are no longer the only ones that get to play pretend.
I don't know anything about you personally, but most "software engineers" are anything but.
I've worked as a software engineer with different types of engineers (electrical, mechanical and automation).
Their testing is often more strict but that is a natural consequence of their products being significantly harder to fix in the field than a software product is.
Other than that, my experience is that our way of working on projects across disciplines is very similar.
It’s all too varied to put people into one or two camps.
If you define "vibe-coders" as people who just write prompts and don't look at code - no, they ain't coders now.
But if you mean people who do LLM-assistet coding, but still read code (like all of those who are upset by this change) - then sure, they always have been coders.
If I got messages like "Accessed 6 websites" I'd flip and go spam a couple github issues with as much "I want names" as I could.
I exactly do not need a "verbose" mode, that lost all value to me as a replacement for something it still is no good at replacing.
You actually argue, that I do not loose anything, when in fact your product just got made worse in two significant areas. And you keep arguing, that shooting the product into one foot is solved by shooting the other foot. Sorry. Not working for me.
Will be evaluating your competition. Was on the cusp of upgrading max to the higher tier. Now? No chance of that happening.
Not like Epstein at all then.
The problem is if you're interdisciplinary, translating something from one field to one typically considered quite distant, you may not always be aware of historic context that is about to fuck you. Not without deeper insight into what the LLM is choosing to do or read and your ability to infer how expected the behavior you're about to see is.
For what it's worth, it absolutely can, just not when invoked in interactive mode.
(This doesn't really contradict your overall point though.)
(That said I do like being able to SSH in and run an agent that way. But there are other remote access modalities.)
Sam wants money. Dario wants to be your dad.
I'm going with Sam.
They are much more responsive on GitHub issues than Anthropic so you could also try reporting your issue there
Electron? The tech that is literally incapable of rendering large amounts of anything, including text, quickly?
This is a names vs. structure thing. For a moment, taboo the term product manager.
What I'm suggesting is a low risk way to see if an engineer has an aptitude for aligning the roadmap with what the users want. If they aren't great at it, they can go back to engineering. We also know for sure that they are technically competent since they are currently working as an engineer, no risk there.
The conventional wisdom (bad meme) is going to the labor market with a search term for people who claim to know what the users want, any user, any problem, doesn't matter. These people are usually incompetent and have never written software. Then hiring 1 and potentially more of the people that respond to the shibboleth.
If you want the first case, then you can't say "product manager" because people will automatically do the second case.
It doesn't have to be the most socially competent engineer to gather feedback. Having the engineering team sit with the target users gives so much insight into how the product is being used.
I once worked on an administrative tool at a financial institution. There were lots of pain points, as it started as a dev tool that turned into a monstrosity for the support staff. We asked to have a meeting with some reps who were literally 2 floors below us. Having the reps talk as they worked with the tool in real time over 1 hour was worth more than a year's worth of feedback that trickled in. It's one thing to solicit feedback. It's another to see how idiosyncrasies shape how products get used.
I agree your assessment about the value of good PMs. The issue, in my experience, is that only about 20% (at most) are actually good. 60% are fine and can be successful with the right Design and Engingeering partners. And 20% should just be replaced by AI now so we can put the proper guardrails around their opinions and not be misled by their charisma or whatever other human traits enabled them to get hired into a job they are utterly unqualified for.
That good taste doesn't translate between domains very often. Good taste for developer tools doesn't mean good taste for a video game inventory screen. And that's the crux of the problem. There is a segment of the labor market calling themselves "product manager" who act like good taste is domain independent, and spread lies about their importance to the success of every business. What's worse is that otherwise smart people (founders, executives) fall for it because they think hiring them is what they are supposed to do.
Over time, as more and more people realized that PM is a side door into big companies with lots of money, "Product Manager" became an imposter role like "Scrum Master". Now product orgs are pretty much synonymous with incompetence.
Taste on the other hand is about creating an overall feeling from a product. It's holistic and about coherence, where intuition is more bottom-up problem solving. Tasteful decisions are those that use restraint, that strike a particular tone, that say 'no' when others might say 'yes'. It's a lot more magical, and a lot rarer.
Both taste and intuition are ultimately about judgment, which is why they're often confused for one another. The difference is they approach problems from the opposite side; taste from above, intuition from below.
I agree with your assessment otherwise, PM can be a real smoke screen especially across domain and company stage.
That’s definitely one of the biggest problems with product management. The delusion that you can be an expert at generic “product”.
We used to have subject matter experts who worked with engineers. That made sense to me.
The counter to that is "the proportion of 'really good engineers' to product engineering teams has got to be in the single digits," and I would agree with that, as well.
The problem is what is incentivized to be built - most teams are working on "number go up?" revenue or engagement as a proxy to revenue "problems." Not "is this a good product that people actively enjoy using?" problems.
Just your typical late-stage capitalism shit.
If people find it too noisy, they can use the flag or toggle that makes everything quieter.
p.s. Serendipitously I just finished my on-site at anthropic today, hi :)
Do you guys have a screen reader user on the dev team?
Is verbose mode the same as the old mode, where only file paths are spoken? Or does it have other text in it? Because I tried to articulate, and may have failed. More text is usually bad for me. It must be consumed linearly. I need specific text.
Quality over quantity
Honestly, man, this is just weird new tech. We're asking a probabilistic model to generate English and JSON and Bash at the same time in an inherently mutable environment and then Anthropic also release one or two updates most workdays that contain tweaks to the system prompt and new feature flags that are being flipped every which way. I don't think you have to believe in a conspiracy theory to understand why it's a little wobbly sometimes.
If Anthropic do something you don't like, you just set a few environment variables and suddenly you're using the Claude Code harness with a local model, or one of thousands available through OpenRouter. And then there is also OpenCode. I haven't tried this, but I'm not worried.
^ https://github.com/ruvnet/claude-flow/wiki/Using-Claude-Code...
This "emdash" and "double dash" discussion and mention is the first time I have heard of it or seen discussion of it. I've never encountered it in the wild, nor seen it used in any meaningful way in all my time on the internet these last 27 years.
And yes - I've seen that special dash character in word for many years. Not once has anyone said "oh hey I type double dashes and word uses that". No it's always been "word has this weird dash and if you copy-paste it it's weird", and no one knows how it pops up in word, etc.
And yes, I've seen the AI spit out the special dash many times. It's a telltale sign of using LLM generated text.
And now, magically, in this single thread, you can see half-dozen different users all using this "--" as if it's normal. It's like upside down world. Either everyone is now using this brand new form of speaking, or they're covering for this Claude code developer.
So yeah, maybe I've been sticking my head in the sand for years now, or maybe I just blindly ignored double-dashes when reading text till now. But it sure seems fishy.
Here are my pre-2020 HN comments, with 3 double hyphens in 8 comments: https://hn.algolia.com/?dateEnd=1576108800&dateRange=custom&...
As I was in the process of typing the search term to get my comments (and had just typed 'author'), this happened to come up as the top search result for Comments by Date for Feb 1st 2000 > Dec 12th 2019: https://news.ycombinator.com/item?id=21768030
Note that I wasn't searching directly for the double hyphen, which doesn't seem to work -- the first result just happened to contain one. If I'm covering for the Anthropic guy, I could be lying about the process by which I found that comment, but I think you should at least see this as sufficient reason to question your assumptions and do some searches of your own.
Not just because it requires constant attention which will eventually lapse, but because the agent has an unlimited number of ways to exfiltrate the key, for example it can pretend to write and run a "test" which reads your key, sends it to the attacker and you'll have no idea it's happening.
It's not.
Engineers are having more and more minutia and busy work taken off their plate, now done by AI. That allows them to be heads up more often, more of their cognitive capacity is directed towards strategy, design, quality.
Meanwhile, users are building more and more of their own tools in house. Why pay someone when you can vibe code a working solution in a few minutes?
So product managers are getting squeezed out by smarter people below them moving into their cognitive space and being better at solving the problems they were supposed to be solving. And users moving into their space by taking low hanging fruit away from them. No more month long discussions about where to put the chart and what color it should be. The user made their own dashboard and it calls into the API. What API? The one the PM doesn't understand and a single engineer maintains with the help of several LLMs.
If it's simple and easy: the user took it over, if it's complex: it's going to the smartest person in the room. That has never been the PM.
Yet the PM always has the last say on what goes in the product, NOT the engineer. Funny how that works...
None of your conclusions are consistent with experience (interviewed 900+ SaaS management teams)
This also depends on your definition on "smartest".
Now this is a good, thoughtful response! Totally agree that if you can convey more information using basically the same amount of space, that's likely a better solution regardless of who's using the product.
100% this.
It might be convenient to hide information from non-technical users; but software engineers need to know what is happening. If it is not visible by default, it should be configurable via dotfiles.
Nope! Not what I said. I specifically said that I don't know if Anthropic is using the information they have well. Please at least have the courtesy not to misrepresent what I'm saying. There's plenty of room to criticize without doing that.
> It's also good to have the humility to know that your subjective opinion as someone not in the target audience even if you're designing the product is less informed in many ways than that of your users.
Ah, but you don't know I'm not the target audience. Claude Code is increasingly seeing non-developer users, and perhaps Anthropic has made a strategic decision to make the product friendlier to them, because they see that as a larger userbase to target?
I agree that it's important to have humility. Here's mine: I don't know why Anthropic made this decision. I know they have much more information than me about the product usage, its roadmap and their overall business strategy.
I understand that you may not like what they're doing here and that the lack of information creeps you out. That's totally valid. My point isn't that you're wrong to have that opinion, it's that folks here are wrong to assume that Anthropic made this decision because they don't understand what they're doing.
I'll just reiterate my initial point that the author of the post and the people commenting here have no idea what information Anthropic is working with. I'm not saying they've made the right decision, but I am saying that people ought to give them the slightest bit of credit here instead of treating them like idiots.
I just want useful tools.
There has been two or three instances that I can remember when it did not block YouTube ads correctly for a couple of days. But those were quickly patched and it started to work again.
One day I visited DistroWatch.com. The site deliberately tweaked its images so ad blockers would block some "good" images. It took me awhile to figure out what was going on. The site freely admitted what it was doing. The site's point was: you're looking at my site, which I provide for free, yet you block the thing that lets me pay for the site?
I stopped using ad blockers after that. If a site has content worth paying for, I pay. If it is a horrible ad-infested hole, I don't visit it at all. Otherwise, I load ads.
Which overall means I pay for more things and visit less crap things and just visit less things period. Which is good.
not even joking
I do that as well. For me it is almost exclusively the case with the news sites.
> If it is a horrible ad-infested hole, I don't visit it at all.
Same.
> Otherwise, I load ads.
There is no "otherwise" for me. I simply do not want to load any kind of ads or "sponsored" content. I see no reason, either moral, ethical or other, to ever do that.
You mean those "free" games, that are hard and grindy by design and the offered help comes in the shape of payed perks to solve the challenges?
Which the parent conveniently left out a definition of. I sort of ignored that implication and went straight to the point - which is it doesn't matter if the PM is the smartest or not. What matters is who makes the decisions and typically the PM does.
Yeah but like I said, people make mistakes. The thing about text output is that it's impossible to track if people are using it at first. You can measure button clicks and key presses. You can't measure eye gaze (at least not usually!).
The good news is, if you remove it and get complaints, you can measure complaints. If you put in a toggle to re-enable it, you can measure how many people activate the toggle. Then you actually have the data, so you can decide whether to just bring it back entirely, or keep it as a toggle, or what.
PM's and designers aren't omniscient. If a feature is view-only, you literally can't tell how much it's used, and it might be minor enough that you never ask about it in user interviews.
This is a common failure complain of all large companies that use metrics but don't actually talk to people. It's not a new complaint at all. And at the day it's one only solved by users throwing an apocalyptic fucking fit at the group producing the software.
>If a feature is view-only, you literally can't tell how much it's used
Then think fucking twice and don't touch it.
Large companies do talk to people. User interviews are extremely common -- it's standard practice that you need to use metrics and talk to users. But like I said, interviews aren't going to cover every tiny little detail.
PM's and designers usually do think twice. But they're human, they're not omniscient. So maybe show a little grace?
For instance, opencode has /undo command which allows you to scroll back and edit a prior prompt. It also support forking conversations based on any prior message.
I think it depends on the set up. I overwrote the default planning agent prompt of opencode to fit my own use cases and my own mcp servers. I’ve been using OpenAI’s gpt codex models and they have been performing very well and I am able to make it do exactly what I ask it to do.
Claude code may do stuff fast, but in terms of quality and the ability to edit only what I want it to do, I don’t think it’s the best. Claude code often take shortcuts or do extra stuff that I didn’t ask.
Anthropic's Opus 4.6 is a bit bigger, but they'd have to be insanely incompetent to not make a profit on inference.
[1] https://github.com/deepseek-ai/open-infra-index/blob/main/20...
An ad-blocker /is/ security software. You don’t have to take it from me, you can read from the Cybersecurity and Infrastructure Security Agency
> AT-A-GLANCE RECOMMENDATIONS
> Standardize and Secure Web Browsers
> Deploy Advertisement Blocking Software
> Isolate Web Browsers from Operating Systems
> Implement Protective Domain Name System Technologies
Literally their second recommendation on this pamphlet about securing web browsers: https://www.cisa.gov/sites/default/files/publications/Capaci...
Moreover you don’t even need a 0-day to fall for phishing. All you need is to be a little tired or somehow not paying attention (inb4 “it will never happen to ME, I am too smart for that”)
Modern advertisement is malware.
Having said that, I do think that there is an investment bubble in AI, but am just arguing that you're not looking at the right signal.
The user is highly skilled; let them filter out what is important
This should be better than adding an indeterminate number of toggles and settings, no?
verbose i think puts it on the TUI and i cant particularly grep or sed on the TUI
The real annoying thing about Opus 4.5 is that it's impossible to publicly say "Opus 4.5 is an order of magnitude better than coding LLMs released just months before it" without sounding like a AI hype booster clickbaiting, but it's the counterintuitive truth, to my personal frustration.
I have been trying to break this damn model since its November release by giving it complex and seemingly impossible coding tasks but this asshole keeps doing them correctly. GPT-5.3-Codex has been the same relative to GPT-5.2-Codex, which just makes me even more frustrated.
CC confidently iterated until it discovered the issue. CC confidently communicated exactly what the bug was, a detailed step-by-step deep dive into all the sections of the code that contributed to it. CC confidently suggested a fix that it then implemented. CC declared victory after 10 minutes!
The bug was still there.
I’m willing to admit I might be “holding it wrong”. I’ve had some successes and failures.
It’s all very impressive, but I still have yet to see how people are consistently getting CC to work for hours on end to produce good work. That still feels far fetched to me.
It strains belief that anyone working on a moderate to large project would not have hit the edge cases and issues. Every other day I discover and have to fix a bug that was introduced by Claude/Codex previously (something implement just slightly incorrect or with just a slightly wrong expectation).
Every engineer I know working "mid-to-hard" problems (FANG and FANG adjacent) has broken every LLM including Opus 4.6, Gemini 3 Pro, and GPT-5.2-Codex on routine tasks. Granted the models have a very high success rate nowadays but they fail in strange ways and if you're well versed in your domain, these are easy to spot.
Granted I guess if you're just saying "build this" and using "it runs and looks fine" as the benchmark then OK.
All this is not to say Opus 4.5/6 are bad, not by a long shot, but your statement is difficult to parse as someone who's been coding a very long time and uses these agents daily. They're awesome but myopic.
You might argue I'm No True Engineer because these aren't serious projects but I'd argue most successful uses of agentic coding aren't by FANG coders.
Also, "order of magnitude better" is such plainly obvious exaggeration it does call your objectivity into question about Opus 4.5 vs. previous models and/or the competition.
I am someone who has spent a lot of time with Sonnet 4.5 before that and was a very outspoken skeptic of agentic coding (https://news.ycombinator.com/item?id=43897320) until I gave Opus 4.5 a fair shake.
But I have to give it to Amodei and his goons in the media, their marketing is top notch. Fear-mongering targeted to normies about the model knowing it is being evaluated and other sort of preaching to the developers.
Metrics definitely lie, but generally in a different way to users/others. It's important to not let the metric become the goal, which is what often happens in a metric-heavy environment (certainly Google & FB, not sure about the rest of big tech).
Autocorrect hall of famer, there.
Are you aware of how many years Amazon didn’t turn a profit?
Not agreeing with the tactic - just…are you aware of it?
Anyway, this has nothing to do with whether inference is profitable.
vs give away the farm via free tier accounts, free trials,
and last but not least: subsidized compute to hook entire organizations………
And how are 7 other providers able to offer DeepSeek API access at roughly the same price as DeepSeek?
This is the expected, the normal, why are you so defensive?
I think you and I have different definitions of “one-shotting”. If the model has to be steered, I don’t consider that a one-shot.
And you clearly “broke” the model a few times based on your prompt log where the model was unable to solve the problem given with the spec.
Honestly, your experience in these repos matches my daily experience with these models almost exactly.
I want to see good/interesting work where the model is going off and doing its thing for multiple hours without supervision.
I'd be hesitant to use that as a way to evaluate things. Different systems run at different speeds. I want to see how much it can get done before it breaks, in different scenarios.
> And you clearly “broke” the model a few times based on your prompt log where the model was unable to solve the problem given with the spec.
That's less due to the model being wrong and more due to myself not knowing what I wanted because I am definitely not a UI/UX person. See my reply in the sibling thread.
> This crate was developed with the assistance of Claude Opus 4.5 initially to answer the shower thought "would the Braille Unicode trick work to visually simulate complex ball physics in a terminal?" Opus 4.5 one-shot the problem, so I decided to further experiment to make it more fun and colorful.
Also, yes, I don’t dispute that human written software takes iteration as well. My point is that the significance of autonomous agentic coding feels exaggerated if I’m holding the LLM’s hand more than I have to hold a senior engineer’s hand.
That doesn’t mean the tech isn’t valuable. The claims just feel over exaggerated.
Let me get this straight: in the Bible, the scapegoat does survive, while the "pure" goat that did nothing wrong gets killed? That's... messed up, even for a tribal rite.
It's not 10x by any means but it doesn't need to be at most dev salaries to pay for itself. 1.5x alone is probably enough of an improvement for most >jr developers for a company to justify $1000/month.
I suppose if your area of responsibility wasn't very broad the value would decrease pretty quickly so maybe less value for people at very large companies?
Using Claude Code for one year is worth the same as a used sedan (I.E., ~$12,000) to you?
You could be investing that money!
edit: Fuck I'm getting trolled
I think personal subs are subsidized while corporate ones definitely not. I have CC for my personal projects running 16h a day with multiple instances, but work CC still racks way higher bills with less usage. If I had to guess my work CC is using 4x as little for 5x the cost so at least 20x difference.
I am not going to say it has 10xed or whatever with my productivity, but I would have never ever in that timeframe built all those things that I have now.
There's also the embarrassing corner physics bugs present in that video, which was something that required a fix in the first few prompts.
And so the very first thing that the LLM does when planning, namely choosing which files to read, are a key point for manual intervention to ensure that the correct domain or business concept is being analyzed.
Speaking personally: Once I know that Claude is looking in the right place, I'm on to the next task - often an entirely different Claude session. But those critical first few seconds, to verify that it's looking in the right place, are entirely different from any other kind of verbosity.
I don't want verbose mode. I want Claude to tell me what it's reading in the first 3 seconds, so I can switch gears without fear it's going to the wrong part of the codebase. By saying that my use case requires verbose mode, you're saying that I need to see massive levels of babysitting-level output (even if less massive than before) to be able to do this.
(To lean into the babysitting analogy, I want Claude to be the babysitter, but I want to make sure the babysitter knows where I left the note before I head out the door.)
To be clear: we re-purposed verbose mode to do exactly what you are asking for. We kept the name "verbose mode", but the behavior is what you want, without the other verbose output.
Might it have been better to retire and/or rename the feature, if the underlying action was very different?
I work on silly basic stuff compared to Claude Code, but I find that I confuse fewer users if I rename a button instead of just changing the underlying effect.
This causes me to have to create new docs, and hopefully triggers affected users to find those docs, when they ask themselves “what happened to that button?”
Verbose mode feels far too verbose to handle that. It’s also very hard to “keep your place” when toggling into verbose mode to see a specific output.
While we're here, another thing that's annoying: the token counter. While claude is working, it read some files, makes an edit, let's say token counter is at 2k tokens, I accept the edit, now it starts counting very fast from 0 to 2k and then shows normal inference speed changes to 2.1k, 2.3k etc. So wanted to confirm: is that just some UI decision and not actually using 2k tokens again? If so, it would be nice to have it off, just continue counting where you left off.
Another thing: is it possible to turn off the words like finagling and similar (I can't remember the spelling of any of them) ?
Big +1 on that. I find the names needlessly distracting. I want to just always say a single thing like “thinking”
"spinnerVerbs": {
"mode": "replace",
"verbs": ["Thinking"]
}
As a counterview, I like the whimsical verbs. I'll be sticking with them. But nice to see there is an option.
All we want is the file paths. That is all. Verbose mode pulls in a lot of other information that might very well be needed in other contexts. People who want that info should use verbose mode. All we want is the regular non-verbose mode, with paths.
I fail to see how it is confusing to users, even new users, to print which paths were accessed. I fail to see the point of printing that some paths were accessed, but not which.
It's not an easy UI problem to solve in all cases since behavior in CC can be so flexible, compaction, forking, etc. But it would be great if it was simply consistent (ctrl+o shows last N where N is like, 50, or 100), with ctrl+e revealing the rest.
“Did something 2 times”
That may as well not be shown at all in default mode?
What useful information is imparted by “Read 4 files”?
You have two issues here:
1) making verbose mode better. Sure.
2) logging useless information in default.
If you're not imparting any useful information, claude may as well just show a spinner.
Like, I'm open to the idea that I'm the one using your software the wrong way, since obviously you know more about it than I do. What would you recommend I do with the knowledge of how many files Claude has read? Is there a situation where this number can tell me whether the model is on the right track?
I actually miss being able to see all of the thinking, for example, because I could tell more quickly when the model was making a wrong assumption and intervene.
I do not know how this concept would work in these agentic environments, but would seem useful, in an environment that has a lot of parallel things going on, with a lot of metrics that could be useful, you would want to have multiple monitors that can be quickly customized with standard linux utilities. Token usage, critical directory access, etc.
This, in conjunction with a config file to define/filter out the log stream should be all that's needed to provide as much or as little detail that would be needed to monitor how things are going, and to alert when certain things are going off the rails.
I find this decision weird due to claude _code_, while being used by _some_ non-technical users, is mostly used by technical users and developers.
Not sure why the choice would be to dumb the output down for technical users/developers.
The thinking mode is super-useful to me as I _often_ saw the model "think" differently from the response. Stuff like "I can see that I need to look for x, y, z to full understand the problem" and then proceeds to just not do that.
This is helpful as I can interrupt the process and guide it to actually do this. With the thinking-output hidden, I have lost this avenue for intervention.
I also want to see what files it reads, but not necessarily the output - I know most of the files that'll be relevant, I just want to see it's not totally off base.
Tl;dr: I would _love_ to have verbose mode be split into two modes: Just thinking and Thinking+Full agent/file output.
---
I'm happy to work in verbose mode. I get many people are probably fine with the standard minimal mode. But at least in my code base, on my projects, I still need to perform a decent amount of handholding through guidance, the model is not working for me the way you describe it working for you.
All I need is a few tools to help me intervene earlier to make claude-code work _much_ better for me. Right now I feel I'm fighting the system frequently.
Or have ctrl+o cycle between "Info, Verbose, Trace"?
Or give us full control over what gets logged through config?
Ideally we would get a new tab where we could pick logging levels on:
- Thoughts
- Files read / written
- Bashes
- Subagents
Because you made stuff up, did not show any proof, and ignored my proof to the contrary.
You made the claim:
> Deepseek lies about costs systematically.
> Meanwhile we have several industry experts pointing not only towards DeepSeek ridiculous claims of efficiency, but also the lies from other labs.
What are those "industry experts" saying that is made up and what is their basis for that?
> You’re citing deepseek unaudited numbers.
Which specific number are you claiming to be fake?
I could just guess blindly and find alternative sources for random numbers from DeepSeek's article.
For example, the tokens-per-second efficiency can also be calculated based on the 30k tps from this NVIDIA article: https://developer.nvidia.com/blog/nvidia-blackwell-delivers-...
But looking for other sources is a waste of my time, when you could just be more precise.
That said, we recently rewrote our renderer to make it much more efficient, so we can bump up the default a bit. Let me see what it feels like to show the last 10-20 messages -- fix incoming.
Write the full content to a file and have less display it. That's a single "render" you do once and write to a file.
Your TUI code spawns `less <file>` and waits. Zero rendering loop overhead, zero tearing, zero stutter. `less` is a 40-year-old tool that exists precisely to solve this problem efficiently.
If you need to stream new content in as the session progresses, write it to the file in the background and the user can use `less +F` (follow mode, like tail -f) to watch updates.
o7
Heck, simply handle the scrolling yourself a la tmux/screen and only update the output at most every 4ms?
It's so trivial, can't you ask your fancy LLM to do it for you? Or you guys lost the plot at his point and forgot the most basics of writing non pessimized code.
Only if you use React as your terminal renderer. You're not rendering 10k objects on screen in a few milliseconds. You're outputting at best a few thousand characters. Even the slowest terminal renderer is capable of doing that.
(And yeah, I would love the verbose mode myself, but there could be various levels to it.)
They did. And the result was a React render loop that takes 16ms to output a hundred characters to screen and tells them it will take a year to rewrite: https://x.com/trq212/status/2014051501786931427
The real problem is their ridiculous "React rendering in the terminal" UI.
(Kidding aside, logging complexity is a slippery slope, and I think it's important, perhaps even at a societal level, for an organization like Anthropic to default to a posture that allows people to feel they have visibility into where their agentic workflows are getting their context from. To the extent that "___ puts you in control" becomes important as rogue agentic behavior is increasingly publicized, it's in keeping with, and arguably critical to, Claude's brand messaging.)