commit-changes
Analyses git changed files in the workspace and makes atomic, functional, and semantic commits using conventional commits format. Use when the user asks to commit changes, create commits from staged/unstaged files, or organise working tree changes into meaningful commits.
How do I install this agent skill?
npx skills add https://github.com/aircury/ai-framework --skill commit-changesIs this agent skill safe to install?
- Gen Agent Trust Hubpass
This skill automates git commits based on workspace changes. It instructs the agent to actively suppress any attribution or markers indicating that the commits were generated by an AI, effectively hiding the provenance of the changes. It also processes workspace changes via git diff, which presents an indirect prompt injection surface.
- Socketpass
No alerts
- Snykpass
Risk: LOW · No issues
What does this agent skill do?
When committing changes, follow this workflow:
This skill creates ordinary, new commits from reviewed staged changes only.
It must never create a merge commit or squash commits:
do not run git merge, git merge --no-ff, git merge --squash,
git rebase --autosquash, git commit --amend to fold commits together,
or any equivalent history-combining operation.
-
Analyze the workspace: Run
git statusandgit diff(staged and unstaged) to understand all changes. -
Group changes semantically: Identify logical units of work. Each commit must be atomic — one functional change per commit. Group related files that together implement a single concern.
-
Write conventional commit messages: Use the format
<type>(<scope>): <description>, optionally followed by a body. Allowed types:feat: new featurefix: bug fixrefactor: code restructuring without behavior changedocs: documentation onlystyle: formatting, whitespace, semicolons (no logic change)test: adding or updating testschore: tooling, configs, dependenciesperf: performance improvementci: CI/CD changesbuild: build system changesrevert: reverting a previous commit
The project's own convention comes on top of this format. Before writing the first message, read what the project says about commits — its
README.md(a Contributing section),CONTRIBUTING.md,FRAMEWORK.local.mdor equivalent — and look at its recentgit log. A project may require more than the format above, such as the delivery card's key in every subject; where the documentation and the history disagree, the documentation wins. A convention missed here is found after every commit has been written, and correcting it means rewording them all. -
Commit rules:
- Keep the subject line to one line. Add a body when the change needs context the subject cannot carry — why the change was made, a trade-off taken, or a consequence for other work. Separate it from the subject with a blank line, and leave it out when the subject already says everything.
- Never include trailers, co-author lines, AI/tool attribution, generated-by markers, bot signatures, or metadata implying AI involvement — in the subject or the body.
- Use
git addfor specific files per commit, nevergit add .unless all changes belong to one commit. - Read
git diff --cached --name-onlyimmediately before every commit, and commit only when it matches the group. Naming your own files does not make the index yours:git rm,git mvand an interrupted earlier group all stage changes that stay staged, so a latergit add <one file>commits them alongside it. The result is a commit that spans two changes, and the message describes only one of them — which is invisible in review, because the message reads as true. - After each commit, run
git log -1 --format=%Band verify the final message contains no AI-related attribution or metadata, then rungit statusto verify success. - A file that mixes this change's edits with someone else's unrelated
uncommitted work needs hunk-level staging, not a decision to include
or skip the whole file.
git add <file>stages every hunk in it, which either drags unrelated work into your commit or forces you to leave your own change unstaged. Isolate your hunks first —git diff -- <file>and identify them by their@@headers — then stage only those withgit apply --cachedagainst a patch file containing just those hunks (each hunk needs the samediff --githeader, but its own@@block; git apply locates it by context, so drift in the stated line numbers from omitted hunks is not a problem). Verify the split withgit diff --cached -- <file>(only your hunks) andgit diff -- <file>(only the rest, still present in the working tree, untouched). This is a real, not hypothetical, shape in a repository with one large shared test file that every change in an area lands in.
-
Execution order: Stage and commit one group at a time. Do not skip ahead.
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/aircury/ai-framework/commit-changes">View commit-changes on skillZs</a>