commit skill

/commit — record a change the way this repo records changes

Optional argument: $ARGUMENTS — treat it as a scope hint or a partial subject the user wants in the message.

Why this skill exists

A commit is not a save button. It is the only durable explanation of a change that travels with the code — the thing someone reads at 2am, six months from now, when git blame lands on a line nobody understands and the ticket system that held the context has since been replaced.

Two failures cause most of the pain, and both are avoidable in the thirty seconds before you type git commit:

• The message narrates the diff. "update user service" tells you nothing the diff does not already say better. The diff is a perfect record of what changed and a useless record of why. Only the message can carry the why, and only the author knows it — for about a day.
• The commit mixes unrelated changes. A commit that fixes a bug and renames a variable and bumps a dependency cannot be reverted, cherry-picked, or bisected. The moment you need any of those, you need them urgently.

Everything below serves those two points.

Live state

Current branch: !git branch --show-current

Working tree:

git status --short

Unstaged:

git diff --stat

Already staged:

git diff --cached --stat

This repo's actual convention — the style oracle. Match it, do not override it:

git log --oneline -20

Full recent messages (do bodies exist? do trailers exist? what do they look like?):

git log -4 --format='%s%n%n%b%n---'

Top-level layout (candidate scopes, if this repo uses scopes):

ls -1

Step 1 — Branch guard

If the current branch is protected — main, master, develop, release, production, prod, or anything listed under protectedBranches in .claude/best-practices-git.json — stop. Do not commit.

Protected branches are shared history. A commit landed straight on one skips review and skips CI, and if it turns out to be wrong, the fix is a public revert rather than a closed pull request nobody had to see.

Offer to create a branch instead and carry the work over — git switch -c <type>/<slug> is non-destructive and brings uncommitted changes with it. Suggest a name from the repo's existing branch convention (git branch -a shows it); if there is none, <type>/<short-slug> matching the commit type you are about to use. Wait for the user before continuing.

Step 2 — Learn the repo's conventions before writing anything

Read the two git log blocks above and answer these, in your head, for this repo:

• Format. Conventional Commits (feat:, fix(scope):)? Bare imperative subjects? A ticket prefix (PROJ-123: …)? Sentence case or lower case?
• Scopes. If the repo uses them, what is the vocabulary? Derive it from the history and the top-level layout, not from your assumptions about how monorepos are usually organized. If existing subjects have no scopes, do not introduce them.
• Granularity. One tight commit per change, or several small logical ones? Stage to match (git add -p when the repo favors small commits).
• Bodies. Do messages carry a body? How long? Do they reference issues?
• Trailers. Does every commit end with something (Co-Authored-By:, Signed-off-by:, a ticket link)? Reproduce exactly what is already there.

A repo's conventions beat this skill's preferences, always. Consistency is worth more than any individual style rule — the history is only searchable if it is uniform. If .claude/best-practices-git.json sets commit.convention, commit.scopes, commit.trailers, or language, those settings win over your inference; they exist for the case where the team is deliberately changing style.

Step 3 — Partition: one commit, one logical change

Group every changed path (staged and unstaged) by what it is for, not by where it lives. Then check:

• Does this changeset contain more than one reason to change? A bug fix plus an unrelated refactor is two commits. A feature plus a formatting sweep is two commits — and the formatting sweep is the one that will bury the feature in review.
• Does it cross a boundary the repo treats as separate (independent packages in a monorepo, unrelated services, vendored code)? If the history shows one scope per commit, keep it that way.

If the working tree holds more than one logical change, show the user the partition and propose an order. Do not silently commit them together, and do not split without telling them — a split changes what lands on the branch.

Step 4 — Safety screen

Look at the file list before staging. Refuse to stage, unless the user explicitly insists after you have named the file:

• .env, .env.*, *.env
• credentials*, service-account*, *-key.json, *.pem, *.key, id_rsa*, *.p12, *.gpg
• Anything whose name or content looks like a token, an API key, or a connection string with a password in it
• Data dumps the user did not mention by name — CSV exports, scrapes, *.sqlite, *.db, dump.*, backup.*
• Binaries over ~1 MB that were not part of the conversation
• Build output and dependency directories (dist/, build/, node_modules/) — if these show up in git status, the .gitignore is wrong; say so.

Say plainly why. "A pushed credential is a leaked credential — rotating it is the only fix, and it stays in the history of every fork and every clone forever" is the whole argument, and it is worth making once rather than arguing about the file.

Step 5 — Stage by name

git add path/one path/two

Never git add -A, git add ., or git add -u. Those stage whatever happens to be in the tree — including the file another agent wrote thirty seconds ago in a parallel session, the debug print you meant to delete, and the local config you never intended to share. Naming the files is the last checkpoint where a human decision is still possible.

Confirm with git status --short after staging, then read the staged diff itself:

git diff --cached

You are looking for what you did not mean to include: leftover console.log, commented-out code, a stray whitespace reflow. Fix those now — after the commit they cost a second commit.

Step 6 — Write the message

<subject — imperative, one line, fits in ~70 characters>

<body — only when the subject cannot carry the meaning>

<trailers the repo already uses>

Subject. Imperative mood: "add", not "added" or "adds" — a commit describes what applying it does. No trailing period. Concrete over categorical: "reject uploads over 10 MB" beats "improve upload validation". Use the format Step 2 inferred.

Body — this is the part that matters. Write one only when there is a why worth keeping, and then write the why, not the what:

• the constraint that forced this shape ("the vendor API rate-limits at 5 rps, so the pool is fixed at 4")
• the incident behind it ("this silently dropped every message with an emoji in the subject")
• the option not taken, and why ("a cache would be faster but goes stale on rename, which is the case that actually breaks users")
• the consequence a future reader would otherwise trip over ("callers must now await this — the sync path was the only reason the tests passed")

Do not restate the diff in prose. If the body reads like a changelog of the files, delete it — you have written the one thing the reader can already see.

Wrap the body at ~72 characters. If $ARGUMENTS was given, work it into the subject naturally rather than pasting it.

Step 7 — Commit

Use a heredoc so newlines and trailers survive intact:

git commit -m "$(cat <<'EOF'
<subject>

<body>

<trailers>
EOF
)"

Hard rules:

• Never --no-verify. Hooks are the team's agreement with itself. A hook you skipped runs in CI anyway, ten minutes later, in front of everyone.
• Never --amend unless the user asked for an amend in this turn. Amending rewrites a commit that may already be pushed and may already be someone else's base.
• Never commit -a. It stages by side effect and defeats Step 5.

Step 8 — When a hook fails

The commit did not happen. This matters: --amend at this point would rewrite the previous commit — someone else's work — and that is the single most common way this goes badly wrong.

1. Read the hook output and fix the underlying problem in the code.
2. Re-stage the fixed files by name.
3. Make a new commit with the same message.

If the hook keeps failing for a reason unrelated to the diff (a missing tool, a broken environment), surface that to the user. Do not disable the hook.

Step 9 — Report

Four lines at most: the commit hash and subject; the branch; anything you left out (out-of-scope files, refused files); one next step if it is obvious. Do not re-summarize the diff — the user just watched you stage it.

Hard rules, in one place

• Never commit on a protected branch.
• Never git add -A / . / -u. Stage by name.
• Never --no-verify, never --amend unasked, never commit -a.
• Never commit a credential, an .env, or a data dump. Naming it and asking is not optional politeness — it is the last line of defense.
• Never invent a convention the repo does not have.