| name | commit |
| description | Create a git commit that matches the repo's existing style. Use when the user says /commit, "commit this", "make a commit", "commit the changes", or similar. Stages only files relevant to the task, splits unrelated work into separate commits, and never pushes. |
commit
Create a focused git commit (or commits) that match the repository's existing conventions.
When to use
Fire on any commit-shaped request:
/commit- "commit this", "commit the changes", "make a commit", "can you commit?"
- Any natural phrasing where the user's intent is to land the current work as a git commit
Do not fire for questions about commits ("what did I commit last?"), for PR creation, or for pushing.
Workflow
Run these in parallel first:
git status
git diff
git diff --staged
git log -n 10 --oneline
Use the output to answer three questions before writing anything:
- What's changed? — from
status and diff. Note untracked files.
- What style does this repo use? — from
log. Match subject length, prefix conventions (Conventional Commits, ticket IDs, emoji, etc.), body format. Do not impose a style the repo doesn't already use.
- Is this one concern or several? — if the diff covers unrelated work, plan multiple commits.
Staging
Stage only files relevant to the task being committed. Skip:
- Unrelated WIP in other files
- Editor/OS cruft (
.DS_Store, swap files) unless intentional
- Local config drift the user didn't ask to commit
Never use git add -A, git add ., or git add -u. Always name files explicitly:
git add path/to/file1 path/to/file2
If the user has already staged something, respect that staging as their intent — don't unstage without asking.
Splitting
If the diff mixes unrelated concerns, make multiple focused commits in sequence. For each commit: stage that concern's files, commit, then move to the next. Announce the split briefly before the first commit (e.g. "Splitting into 2 commits: auth refactor, then unrelated README fix").
Writing the message
- Mirror the repo's existing style from
git log.
- Focus on why, not what — the diff shows what.
- Subject in imperative mood if that's the repo convention.
- Pass the message via HEREDOC to preserve formatting:
git commit -m "$(cat <<'EOF'
Subject line
Optional body explaining why.
EOF
)"
After committing
Run git status to confirm the tree is clean (or shows only the expected remaining unstaged changes if splitting across commits).
Do not push. Pushing is a separate action the user must request.
Handling pre-commit hook failures
If a pre-commit hook fails, the commit did not land. Do:
- Read the hook output. Fix the underlying issue in the code.
- Re-stage the fixed files (
git add <files>).
- Create a new commit — do not use
--amend (there's nothing to amend) and do not use --no-verify to bypass the hook.
If the hook's fix is out of scope or unclear, stop and surface the error to the user.
Guardrails
Stop and confirm with the user before committing if any of these apply:
- Mass-staging attempt — if you're tempted to run
git add -A, git add ., or git add -u, stop. Name files explicitly instead. This is a hard rule from the user's global config.
- Large diff — >500 changed lines or >20 changed files. Summarize what's in it and ask whether to proceed as-is or split.
- Sensitive files in the diff —
.env, .env.*, files named *credentials*, *secret*, *.pem, *.key, id_rsa*, *.p12, or anything with obvious API-key-shaped strings. Warn before including them.
- Lockfile-only or lockfile-dominant commit — if the commit is just
package-lock.json / yarn.lock / Gemfile.lock / mix.lock with no matching manifest change, confirm that's intentional.
Hard rules
- Never
git add -A, git add ., or git add -u.
- Never push.
- Never
--amend a commit you didn't just create in this turn.
- Never
--no-verify to skip hooks.
- Never commit files the user didn't touch in this session without flagging them first.
