The Lost Art of Commit Messages(seyhan.me) |
The Lost Art of Commit Messages(seyhan.me) |
Why do we care about commit messages? I only read them when rebasing.
Besides, a few passes refactoring and your git history is ruined. No way you’ll find the original commit within any reasonable time frame.
If I move the code to another function, your original commit message will be hard to trace. If the code was commented, the comment would have been moved along. If the code was readable, you don’t need the comment or the commit message.
With structured comments, you can even use the AI to fill out the PR template easily and precisely.
Reject commit message suggestions
When working on large software projects, commit messages that follow
the format suggested by the author rarely provide additional value.
Having commits prefixed by "chore:" or "docs(ui):" aren't that useful.
Instead, some of those 50 characters can be used for more descriptive
titles. Commit messages are often the best and only context available
when bisecting a bug, so bullet points only really make sense when
making a list. Prose works just fine.
Writing styles vary across contributors but it's much more useful to
get *anything* from someone in their preferred format than nothing at
all because they are frustrated with rules. Some committers do have a
special "committer voice" that they use when describing complex changes
in a commit message. For example, the text here is slightly abridged
and focuses less on the first person than what is typically expected.
This evolves naturally from writing these.
Finally, commit comments definitely should have jokes in them. This is
actually more critical than wrapping them at 72 characters.This is the downside of the variation in style across contributors. It works if people want to build something, and not just do the minimum. Open source projects frequently have no such limitations.
But I am sad to report that some people appear to need the structure.
Address structural issue with code contributions
In some cases, it's possible to end up with commit messages that carry
very little value. Unfortunately, while the minimum is clearly
insufficient, it is possible to "expand" the minimum into a specific
format without substantially improving the quality of the commit
message. For example, a substantial refactor can easily just be
described as a list of the changes that it included, rather than a
rationale for why it was done (which is not readily apparent from the
change itself).
In the long run, problems involving people can be improved by process,
but cannot wholly be solved by them. Pernicious failures need to be
acted upon directly. Demonstrating the value of good commit messages is
often a better driver of improvements than a template that is a chore
to fill out.One of the easiest ways to start there is to adopt PR-based review systems, where any number of commits can get pushed to a git repository and then reviewed as a single block. This is fairly inflexible, but it is very simple to get started with (no need to learn any special rebasing commands, just add, commit, add, commit, push). Most git forges allow you to do squash merges, so everything in the PR will become a single commit, usually with the PR description as a commit message. Then you can refine the commit and commit message during the review process.
There are better tools out there, but in my experience, this is the quickest way to get started for most teams, especially as you're probably already using a forge that allows you to do all of this already.
Also, reconsider column-based hard-wrapping entirely,
and consider wrapping mostly on punctuation,
or other reasonable opportunities when it gets a bit long
and you don’t really want to insert punctuation.
This is not without its drawbacks and limitations—
most software when rewrapping won’t cope with trailing em dashes,
and you’ll end up with extraneous spaces when the lines are merged,
even though this is *always* how wrapping has worked around dashes.
Over time, this will definitely influence your writing style.
Your sense of æsthetics will push you to reword sentences,
because otherwise the edge is too ragged for your liking,
and at some point you’ll see a pattern in a paragraph,
and feel an uncontrollable urge to continue it,
and you’ll waste a bunch of time tweaking,
and maybe no one will appreciate it.
But you’ll appreciate it.
It’ll be fun.
Mostly.
But seriously, you *will* become more aware of things like clause length,
and your writing will probably improve.
And after writing verbosely for a long time,
you’ll focus on concision,
and that’ll be another challenge.
And it’ll be fun too.
Mostly.
Then probably someone will take over your pet docs project,
and mandate an autoformatter that rewraps it to 59 columns.
Stubbornly, in the quiet of your mind, you’ll make it work.
It’s a greater challenge, but you won’t lack determination.
Style guides come, and style guides go; when this one goes,
your lines can remain the same, needing no fixing, perfect.
—⁂—
And really, just look at how the last paragraph of the parent comment looks,
before and after:
Finally, commit comments definitely should have jokes in them. This is
actually more critical than wrapping them at 72 characters.
—⁂—
Finally, commit comments definitely should have jokes in them.
This is actually more critical than wrapping them at 72 characters.
Much more beautiful, is it not?commit e6f22569e9db16dbef7d9bdee5bb41b8cb7a03ca
Author: Jim Warner <james.warner@comcast.net>
Date: Tue Jul 23 00:00:00 2024 -0500
top: attempt once again to allow help text translation
Back when our release 4.0.1 was being readied, sources
were sent to the TP (Translation Project). However one
person named Benno Schulenberg refused to release them
for translation. His stated reason was the top command
line help text which then finally included long forms.
He demanded that the help text be broken into separate
strings instead of a single large string. But, all the
top text (some much more complex) has just one string.
So that stated reason was, at the least, inconsistent.
[ I suspect the real reason was that Mr. Schulenberg ]
[ thought that the carefully right-justified English ]
[ wording would also be required of translations too ]
The bottom line was that Benno took it upon himself to
change the TP motto from "you code, we translate" into
"first we tell you how to code and then we translate".
Rather than bend a knee to that despot, I disabled the
text entirely, admittedly denying users a translation.
Now, with this commit we enable translatable help text
but with a hint included to ignore the justified text.
Reference)s):
. Oct, 2022 - finalized translation exclusion
commit ab05a3785f29cc4b754e17c53bfb3d8ba054563e
Signed-off-by: Jim Warner <james.warner@comcast.net>I don't use those for humans, I use those for tooling, such as semantic-release. The human-readable bits come after that.
One word naming the topic or area or system that was changed, then colon separated with a very short sentence giving a summary of the changes, then two lines later (if necessary), a bullet point list of the most important/noteworthy changes, then an explanation for why a thing was changed (if any change in the commit warrants it).
Honestly, I know most people won't go beyond the first line, but I do find the rest of it very helpful for my own work if I have to go through the commits sometime in the future.
It also helps that most of it is optional and I decide on a case by case basis whether just the first line is sufficient or I need the whole thing.
I see in the comments people questioning the purpose of a bullet point list, but it actually is helpful. I don't want to have to check the diff for every single commit if I don't have to. It's time consuming. If a commit message can tell me immediately if it touched something I'm interested in, that's a big time and effort and mental bandwidth saver.
Example:
auth: Refactored and fixed edge cases
- Fixed incorrect handling of token groups
- Added role enum to replace static strings
Example commits of something I worked on a few days back:
$ git l feature/character-selection
c54825f 3 days ago Robert Schaap (feature/character-selection) Simplify color picker, fetch current color
d512569 3 days ago Robert Schaap Fix recolor for female, clean up files
6d05ce4 3 days ago Robert Schaap Add color picker to change shirt color
441180b 3 days ago Robert Schaap Show male in editor
17045dc 3 days ago Robert Schaap Remove old character
95772ff 3 days ago Robert Schaap Add characters
$ git show HEAD~1
commit be50e0d701d565cebdf4f29e0c9d8030a1a8faf7
Author: Robert Schaap
Date: Mon Mar 24 21:29:20 2025 +0100
Character selection (#14)
* Add characters
* Remove old character
* Show male in editor
* Add color picker to change shirt color
* Fix recolor for female, clean up files
* Simplify color picker, fetch current colorConsider a trivial change but affecting tens of places in the code, eg an API change. It's very useful to be able to quickly glance past "use new abc-API; required since dependency X bumped to Y" and mentally move on, rather than actually having eyes scanning over those actual changes.
I do not understand why people insist on trowing inane technical solutions at social problems. It doesn't work.
In my experience conventional commits tend to lower character counts and improve the readability of messages. `bug(auth): adjust XYZ` is shorter than `fix auth bug by adjusting XYZ`.
I wouldn't say its a lost art... we do this at our company
[0]: https://www.bekk.christmas/post/2019/11/gitmoji-yay-or-nay [1]: https://news.ycombinator.com/item?id=21760021
There are times when I make a one-line change and write a paragraph or two explaining why it had to be done. But these kinds of things often drown in the noise of a dozen other changes. If that one was important enough, I will reference it in an ongoing discussion or documentation, or at least include "read below:" on the first line.
I usually see people include a more elaborate commentary with the pull request. If the changeset is good but the series of individual commits is a bit messy, just merge by squashing.
(Also: this comment is meta.)
Compare the Linux commit history, every commit has its full context and explanation and they do not rely on external systems.
I"m working on a repository that uses at least four different jira ticket number formats. All commits should have a jira reference but I think only the current format can still be looked up. And maybe the predecessor if you know what jira field to query. All the rest are lost in corporate limbo. Not that those tickets added much more context to the actual commit...
So yeah, always write your commit messages as standalone as possible.
You can not avoid it all the time but maybe It's better to use the PR description for that purpose.
Commit messages are a great place to bury docs nobody will look at.
For example, when typical commits are about solving or implementing some support ticket, issue, new feature etc. that is documented elsewhere correct references are far more important than classification of the commit type or descriptive details; special commits (refactoring, chore, docs etc.) are easily noticeable because they don't reference a specific issue and many details are better omitted for the sake of deduplication.
fix(ui): correct alignment on dashboard widgets
- adjust CSS flex properties
- test on multiple screen sizes
Fixes #204
#204 - adjusted CSS flex properties of dashboard widgets; tested on multiple screen sizes
#204 - same CSS flex properties for all dashboard widget DIV elements; tested on 800*600, 1880*960 and 2900\*1600 browser windows but not on the standard smartphone emulatorsOther peeve: quote the core of the bug report you're closing, so when GitHub inevitably goes away/turns evil/starts charging, you don't lose half your knowledge. The git tree should always stand alone.
On a lighter note, I recommend "8 Types of Commit Messages That Show He's NOT the Man for You" https://web.archive.org/web/20210606005031/https://www.codem...
I'd posit that well-structured commits are principally for the benefit of the reviewer of the code. Order your commits in a fashion that makes sense narratively, and give them meaningful commit messages. Use interactive rebasing liberally if need be to accomplish this goal.
As a reviewer, you are well within your rights to decline a PR that consists of a single non-meaningful commit message and a +/- 1000 lines diff. Effort is expected from both parties in checking in code.
Two usecases I can think of: non-native English speakers and ADHD developers. For those groups, having some tool that autofills the first draft of the commit message (that you can then modify) would probably improve their overall commit message quality.
A quick google gave me several projects:
- https://github.com/Nutlope/aicommits
- https://github.com/insulineru/ai-commit
I'm sure there's more.
Just as you'd expect from an LLM, it hallucinates a description that overfits the average commit message. It doesn't make a good summary of what changed, and most importantly (though I don't expect it to) why it changed, which is what I need in a commit message.
The most important thing to convey in a git commit message is the why, not the what or how.
The title should be a short summary of the change (the what), so that it can easily be searched for.
The description should explain why this change needed to be made. The how and the what can be determined by reading the code. But more often than not, explaining the why helps to clarify intent so that future readers can determine if your rationale still holds.
How is distilling a novel down to a short description and 2 points any better? This format is too limiting.
From the article: "- adjust timeout settings to prevent crashes". Include the details of why the timeout setting lead to crashes; what were the inputs and the cases that caused this.
This lets us decide whether the fix stays or goes the next time there is an issue in the same piece of code, or your commit broke something unrelated - the person fixing it needs to know _why_ you changed the code.
Using a type is to contrived, because your commit will often include e.g documentation and a feature. Splitting them for the sake of following this pattern makes no sense.
The actual purpose is mostly for tooling to help generate release notes.
But it's the only way if there's no commit message discipline, and even if there is it'll be difficult.
On the other hand, job security?
Each commit becomes one "unit of review", and the commit message can be reviewed just like the code changes.
I don't think my team would.
Did the author measure this ? If not, he's the monkey.
Yes. And yet:
> - Keep each point brief and focused.
> - Use bullets (-) and avoid lengthy explanations.
No, they lost me there. You often can’t tell the necessary narrative in a single sentence, and you frequently can’t explain why changes were made briefly, and I strongly refute that “avoiding lengthy explanations” should be a goal in commit messages.
I exceed ten lines very frequently, and when working professionally I’ve tended to exceed fifty lines at least twice a year. Coworkers have commented favourably on my commit messages, as unusually useful, and some have even lengthened theirs more often after experiencing them.
I don’t know what my record is, but it was more than 300, though that was as part of a long-lived refactoring branch that I was needing to rebase every week or two for six months as I continued working on it, migrating from an in-house concatenation-and-#ifdef module builder to ECMAScript Modules, and I was maintaining the dozens of commits meticulously, some regenerated automatically and some requiring manual effort, and so I was detailing the semantics of the changes across a large number of files, so it might not count quite so much. But I’ve done over a hundred a few times for other reasons. Keep your mind open, and look at the longest commit messages in popular repositories, and you’ll find there’s a lot of scope for very long commit messages.
—⁂—
I loathe Conventional Commits. It saps the art out of the commit message, the tags are frequently badly applicable, the benefits are negligible to negative (changelogs shouldn’t be generated from commit logs—they should be their own thing), and they’re simply ugly.
> Add characters
I can probably tell from the code that that's what's happening. But what requirements drove these particular characters?
> Remove old character
What makes it old? How would I recognise an old character in the future?
> Show male in editor
Why did male not show before? Was there a bug or a partially implemented feature?
> Fix recolor for female, clean up files
What does it mean to "fix recolor"? And even worse, what is "clean up files"? What requirements drove this file cleaning? Why were the files unclean in the first place?
etc. Commit messages in the style of "fix X" or "add Y" or "remove Z" or "nondescript action on W" are the bane of my existence. They seem so meaningful but they don't tell me anything when I'm trying to trace why a particular bug was introduced – or whether it's even a bug in the first place.
But a commit message needs to describe what the commit does, when applied. A good rule of thumb - also explained on the git site [0] - is to put it in a template like "When applied, this patch will <your commit message>".
The grandparent comment is almost there though, using the right tense of "fix" instead of "fixed", the latter being in the work log form of "I fixed such and such".
[0] https://git-scm.com/docs/SubmittingPatches/2.2.3#:~:text=Des...
Reducing the human-readability of the most human-presented part of the commit, to make it easier for the can-do-billions-of-operations-a-second machine, seems backwards.