Create commit

Skill rccsousa/skills/skills/create-commit

Stage changes and create a commit with a one-line Conventional Commits message. Groups files by logical change, builds a focused subject, and never adds Claude as co-author. Use when ready to commit, after a focused chunk of work, or when the user says "commit this", "make a commit", "stage and commit".From its SKILL.md

Install

npx -y skills add rccsousa/skills --skill create-commit

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

8.0 KB, as published. Nobody here has run it

Create Commit

Stage + commit w/ Conventional Commits, oneliner only (subject line, no body, no footer). Single-purpose commits, no Claude co-author, hitl by default.

When to use

After a focused chunk of work is ready to land
Inside one-shot implement phase (called by worker agents)
Whenever the user says "commit this" / "make a commit"

Skip if working tree is clean → halt.

Invocation

/create-commit [--all] [--type=<type>] [--scope=<scope>] [--message=<msg>] [--split]
               [--auto --i-understand-the-risk]

Flags:

--all — stage all modified + untracked files (default: detect logical group)
--type=<type> — Conventional Commits type (feat/fix/chore/docs/refactor/test/perf/build/ci/revert)
--scope=<scope> — optional scope
--message=<msg> — override derived subject
--split — propose multiple commits if changes span unrelated logical groups
--auto — skip the confirmation gate (see ⚠️ section below). Requires --i-understand-the-risk as a companion flag. Used by one-shot --mode=auto.

Default behaviour: confirmation gate is ON, every commit requires explicit "go".

Process

Inspect changes. git status --porcelain + git diff --stat + git diff --staged --stat.
Identify logical groups. Heuristics:
- Same top-level dir (e.g. src/auth/* + tests/auth/*) → one group
- package.json + lockfile → one group
- docs/ changes separate from code
- Unrelated dirs → propose split
If multiple groups + --split flag: propose N commits, walk one at a time. Else: bundle all into one commit.
Derive type + scope.
- --type flag → use it
- Else: detect from diff (new file in src/ → feat; bug-fix keyword in diff → fix; only docs → docs; tests-only → test; etc.)
- Scope: top-level dir of changed files
Build subject. Imperative, lowercase, ≤70 chars, no trailing dot. Oneliner only — no body, no footer.
Confirmation gate (skipped only when --auto --i-understand-the-risk both present + gate preconditions hold; see below).
Stage + commit. git add <files> → git commit -m <subject>.
Verify. git log -1 --format="%H %s" → return SHA + subject. In --auto, also append a line to .claude/auto-mode.log.

Confirmation gate

Default ON: surface proposed commit, wait for "go" / "yes" / "commit" before staging

committing.

Format:

proposed commit:

  feat(auth): rotate refresh tokens on use

files (3):
  src/auth/refresh.ts        +42 -8
  src/auth/handler.ts        +12 -3
  tests/auth/refresh_test.ts +88 -0

  [g] go
  [e] edit message
  [s] split (propose multiple commits)
  [c] cancel

⚠️ `--auto` (gate skip)

--auto --i-understand-the-risk lifts the per-commit gate. Used by one-shot --mode=auto to keep the cascade unattended.

Preconditions — all must hold, else refuse + halt:

Both flags present: --auto AND --i-understand-the-risk. No env var, no default. The risk-ack flag has no shorthand.
.claude/auto-mode-disabled not present at repo root.
~/.claude/auto-mode-disabled not present globally.
Current branch != repo default branch (no auto commit on main / master).
No --amend flag (auto + amend is never allowed).
Pre-commit hooks present + would run normally (--no-verify is always refused, even in auto).

If any fails → halt with the specific reason. Do not silently downgrade.

In auto mode:

Surface a one-line audit (auto-commit: feat(auth): ... [3 files]) before staging, but do not wait for input.
Append a line to .claude/auto-mode.log: 2026-05-21T14:32:18Z create-commit <sha> <subject>.
Hard caps still apply: no --no-verify, no --amend, no co-author injection, no identity override. All "Hard rules" below remain in force.

Message format

<type>(<scope>): <subject>

Oneliner only. No body, no footer, no BREAKING CHANGE:, no Closes #N. If a footer feels load-bearing (breaking change, issue link), put it in the PR description instead — the commit stays a single line.

Hard rules (apply always, no override):

Use the caller's git identity (user.name / user.email from local git config). Never set --author, never export GIT_AUTHOR_NAME / GIT_AUTHOR_EMAIL, never inject any tool, plugin, or marketplace-owner identity into the commit.
Never add Co-Authored-By: Claude (or any Claude variant)
Never add 🤖 Generated with Claude Code (or similar AI-source markers)
Never pass -m twice — single subject line only
Never use --no-verify (don't bypass hooks)
Never use --amend unless user explicitly asks (creates a new commit instead)

Type detection heuristics

Diff signal	Type
New file in `src/` / `lib/` / `app/`	feat
`fix`, `bug`, `crash`, `regression` keyword in changes	fix
Only files in `docs/` / `*.md`	docs
Only files in `tests/` / `_test.` / `.test.`	test
Only formatting / import-order changes	chore
Function renamed without behaviour change	refactor
Bench / latency / perf keyword in changes	perf
`Dockerfile`, `package.json` deps, build config	build
`.github/workflows/`, CI config	ci

Multiple signals → pick the most specific. Ambiguous → ask user.

Split detection

Propose --split when changes touch unrelated areas:

src/auth/ + src/billing/ (no shared call site)
src/ + docs/ (mixing code + doc updates)
bug fix + unrelated refactor in same diff

Propose order: smallest blast-radius first (docs → tests → fix → feat → refactor).

User accepts → walk one commit at a time, with the confirmation gate per commit.

Halt conditions

Working tree clean → halt: "nothing to commit"
On detached HEAD → halt: "checkout a branch first"
Pre-commit hook fails → halt w/ hook output; do NOT bypass w/ --no-verify
User cancels at gate → halt cleanly, leave staging area as-is

Output style — caveman ultra (under-the-hood)

Applies to: commit subject, surface text to user during gate.

Does NOT apply to: file paths, identifiers, error messages, SHAs.

Rules:

Subject: imperative verb + object, fragments OK, ≤70 chars
Drop articles, filler
Short synonyms (DB / auth / fn / req / res / impl)

Examples:

Normal: feat(auth): add a check to validate refresh token timestamps within a 5-minute window
Caveman ultra: feat(auth): bound refresh token replay window to 5min
Normal: fix(api): the error in the handler was being silently swallowed
Caveman ultra: fix(api): rethrow swallowed err in handler

Out of scope

Pushing the commit → git push (separate step; /create-pr handles it)
Opening a PR → /create-pr
Amending a previous commit → user-driven only, never automatic
Force-push, rebase, merge → out of scope

Why this skill exists

Conventional Commits + no co-author + hitl confirmation are repeated invariants across the flow. Centralising them here means /one-shot, /code-fix, and worker agents all produce uniform commits without re-implementing the rules.

The confirmation gate defaults to ON. The --auto escape exists for one-shot --mode=auto, gated behind an explicit risk-ack flag and the repo/global auto-mode-disabled veto. The invoker owns the call to lift the gate. Hard caps — no --no-verify, no --amend, no co-author injection — apply in every mode.