I’ll typically have a bunch of short sessions over the course of a day. Anytime I start a task that isn’t going to very directly benefit from the existing context I start fresh.
I don’t find a lot of benefit in explaining the project overall to Claude — I’ve deleted a lot of that explanation from my Claude.md because it didn’t seem to impact much.
I typically start a task by pointing it to 1-2 files and giving it some explanation of what I want done, and it figures it out.
Basically never hit context window limits or compactions, and can’t remember the last time I hit a 5 hour or a weekly limit.
They aren't really "explaining the project" either, but more module- or task-specific preferences, hand reference pointers, or other things like "there are mixed examples of how to do certain things in this project, prefer X to Y." I use a write-everything-twice approach. After I find myself having to correct an implementation because it didn't figure out one of these things on its own from the existing code, I'll add an entry. That also avoids bloating things with "I think this is relevant" compared to "I have noticed that this is necessary."
I keep doing this because it lets me experiment with different approaches to problems without risk of it fixating on things from a previous abandoned attempt, and particularly because sometimes I'm wrong and I haven't found the agent harnesses particularly reliable at taking my word for it from a POV of "yes I know I said we need xyz earlier, but let's please entirely forget about that."
I guess I need to do some claude.md work or find other ways to prime the session so i get the good personality and not the evil twin.
I've been running Claude and GPT in my own agent harness. The main difference I notice is that tasks take about 7x longer to complete if they're run in the official Claude or Codex harness (and cost me 7x more).
You would think this would lead to increased correctness, but that doesn't seem to be the case. Today I tested both side by side. They both resulted in data loss. (I had a backup obviously.)
GPT running in the official harness did a bunch of extra tests and double checking, and ended up with the same result regardless (it permanently deleted a bunch of documentation).
All else being equal, I like getting my data loss 7x faster and cheaper ;)
Normally I do it exactly as you say, point at a few files, but if I know these features are involved I point at the corresponding mds instead. Its a shortcut for me to type less.
It basically takes care of itself, or at least as close as it can.
I can’t imagine this being worth optimizing. The issue is never that Claude can’t figure out what the projects is about…
Am I missing something or does this project not solve a problem most regular people have?
A good method seems to be only make a skill or memory when the LLM gets something wrong, or if you actually observe it's always doing the same step and you can get the model to the same place with less tokens.
The solution that I've developed is, let the agent figure things out efficiently, without inflating the context. I have what I call a smart repo that better explains this at
https://github.com/gitsense/smart-ripgrep
The basic idea is, when the agent does a ripgrep it gets back files + matching lines + context.
Even the /handoff skill was written by the model…
I do think that this project is interesting in several ways - prioritizing privacy, minimizing spend, and using objective semantic markers to sift and consolidate the key takeaways from long sessions. I'd like to try it on my cline project history. But while it would make a great recording of project history, I wonder if a lot of it doesn't end up detailing blind alleys the project went down and had to back out of.
Generally when this happens I feel that it's due to vague specification on my part, or avoiding architectural decisions I didn't want to deal with and implicitly inviting the model to implement a lowest-common-denominator solution.
Not sure I’d call that “stopping wasting my tokens”.
The only tip I can give is that your skill that builds or wraps up work. You should have it update those files if anything has changed.
Claude/Agents files shouldn't be bloated, but should imho act as a basic amount of context on the project so your agent and skills can pick up and go, with even the most basic initial prompt.
Depending on the scope of work you’re doing, it might be better to have this removed from the context of the work that was done.
I keep a “Last Updated Hash” in my md and every so often will have the LLM pull a diff from that hash to the current head, then determine what doesn’t match.
I had the idea of an oracle/apprentice. The idea was that the new session would learn from the old session, like a tutor.
I asked Claude to code a program, so that the old session would launch in loop, being the "oracle", and the new session would use it to connect to the oracle, to ask it questions.
So when I'm doing that, I'm seeing the two Claude sessions discussing between them. It's fun seeing the "apprentice" asking follow-ups questions.
The mechanism (on a Linux machine) is relatively simple : I asked Claude to code a Go tool, to use an abstract socket (\0claude-handoff-<projectname>), and specified that whichever is the first to successfully listen (no EADDRINUSE) becomes the "listen()er/accept()er" and that the second becomes the "connect()er".
So that establishing the socket in whichever order is independent of which of them is the oracle/listener.
I've put the mechanism in a global Claude rules. In the oracle when I'm a 98% usage of the 1M context, I just have to type "handoff <projectname> oracle", and to start a new session with "handoff <projectname> client".
And the "oracle" will loop on the tool (with a subagent, waiting indefinitely), the tool exits with a question on output, and re-call (with a subagent "handoff <projectname> answer") to give back the answer (which automatically waits for the next question).
And since the oracle is doing the call to the handoff tool in a subagent, when you see it answering, you can also type something along the lines of "hey please also precise to the apprentice <some specific information>".
The "transmission of knowledge which matters" is so much efficient, that ~2% of remaining context (20k tokens) is enough to transmis WAY MORE USEFUL information that any memory saving which would miss important informations.
It's not unexpected. It's like real-life. You may have an human put all informations that you want in some documentation, nothing can replace a phone line from the new human to the previous human for specific follow-up questions.
Tho one of the mattering rule specificity is to precise that the oracle should always include in this response the level of confidence in the answer, like if it is certainty/guess/hypothesis.
It's fun seeing the two Claudes discussing like two colleagues. I guess you could also ask Claude to code the tool to instead connect on a localhost IRC server.
I think that if you want this tool, you just have to c/c my text into a new Claude session and to tell it "I want this too, please code it, please also setup the global rule".
And keeping a local copy of everything you ever told Claude in your context window is bad for the same reasons keeping a local copy of your code called My_Code_v3_final.zip is bad.
But if I may, the need to manually update the context is a huge hurdle.
Automation like this is limited unless no human has to remember it. So perhaps you can save context during the PreCompact and Stop hooks.
I saw /graphify recently which cuts down on exploration cost and seems more appealing (although I haven’t tried it yet)
My advice: the best claude is the raw claude, with some custom tailored skills. That’s it, no plugins.
I also imagine that varies by model.
It's slightly bigger now, but here's a ~50 line version for reference. I added the missing outer while-loop, so it takes user input etc.
https://gist.github.com/a-n-d-a-i/bd50aaa4bdb15f9a4cc8176ee3...
I mostly use it with GLM via their coding plan, I got a year for like $20 when it was on sale. But I also hooked it up to Sonnet, Opus, GPT, etc.