atomic-git-commits - SKILL.md Agent Skill

name: atomic-git-commits description: "Write atomic, human-reviewable git commits — one logical change per commit with clear conventional messages" version: 1.0.4

Atomic Git Commits

Write atomic, human-reviewable git commits. Each commit should represent one logical change that can be understood, reviewed, and reverted independently.

What Makes a Commit Atomic

An atomic commit:

Contains one logical change — not "fix bug and also refactor and update deps"
Builds successfully at every point in history — no broken intermediate states
Can be reverted independently without side effects on unrelated code
Has a clear, descriptive message that explains the why, not just the what

Commit Message Format

Use Conventional Commits with scope:

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

<body — optional, explains WHY>

Co-Authored-By: Name <email>

Types

Type	When to Use
`feat`	New feature or capability
`fix`	Bug fix
`refactor`	Code restructuring without behavior change
`docs`	Documentation only
`test`	Adding or updating tests
`chore`	Build, CI, dependency updates
`perf`	Performance improvement

Subject Line Rules

Imperative mood: "add feature" not "added feature" or "adds feature"
Under 70 characters — details go in the body
Lowercase first word after the colon
No trailing period

Examples

feat(web): add session timeout with configurable TTL

refactor(web): replace stateless logger with wide event Logger class

fix(cli): handle 429 rate limit with exponential backoff

docs: add server functions middleware documentation

Splitting Work into Atomic Commits

Strategy: Layer by Dependency

When refactoring, commit in dependency order so each commit compiles:

1. feat(shared): add new validation types        ← foundation
2. refactor(web): update middleware to use types  ← consumer
3. feat(web): enrich auth middleware with logger  ← integration

Strategy: Separate Mechanism from Policy

1. refactor(web): extract Logger class            ← mechanism (the tool)
2. refactor(web): wire Logger into middleware      ← policy (how it's used)
3. feat(web): add user identity to wide events    ← enrichment

Strategy: Infrastructure then Features

1. chore(web): add rate limiting middleware       ← infrastructure
2. feat(api): apply rate limits to publish route  ← feature using it

What NOT to Do

Don't mix concerns

# Bad — two unrelated changes in one commit
git commit -m "fix login bug and update README"

# Good — separate commits
git commit -m "fix(auth): validate email before session create"
git commit -m "docs: update README with auth setup instructions"

Don't commit broken intermediate states

# Bad — breaks the build between commits
commit 1: "remove old auth module"       ← build breaks here
commit 2: "add new auth module"          ← build works again

# Good — atomic swap
commit 1: "refactor(auth): replace old auth with better-auth"

Don't commit generated files with source changes

# Bad — noise in the diff
commit: "feat(web): add dashboard route"
  + src/routes/dashboard.tsx          ← your code
  + src/routeTree.gen.ts              ← auto-generated noise

# Good — separate or .gitignore generated files
commit: "feat(web): add dashboard route"
  + src/routes/dashboard.tsx

Staging Specific Files

Always stage explicitly — avoid git add -A which can catch secrets or generated files:

# Stage specific files for this logical change
git add src/lib/logger.ts src/lib/middleware/types.ts

# Commit with heredoc for multi-line messages
git commit -m "$(cat <<'EOF'
refactor(web): replace stateless logger with wide event Logger class

Introduce a Logger class that accumulates context throughout a request
lifecycle and flushes a single canonical log line.

Co-Authored-By: Name <email>
EOF
)"

Review Checklist

Before each commit, verify:

git diff --staged shows only changes related to this commit's purpose
The commit message accurately describes the change
The project builds after this commit
No unrelated formatting, whitespace, or import changes snuck in
No secrets, .env files, or large binaries are staged
The commit can be understood in isolation during code review

Interactive Rebase for Cleanup

If you made messy WIP commits, clean up before pushing:

# Squash/reorder last N commits
git rebase -i HEAD~N

# In the editor:
pick abc1234 feat(web): add Logger class
squash def5678 wip: fix typo in Logger     ← squash into previous
pick ghi9012 refactor(web): wire Logger into middleware
drop jkl3456 debug: add console.log        ← drop entirely

Important: Only rebase commits that haven't been pushed. Never rebase shared history.

Practical Example: Refactoring a Logger

Given a task to refactor a stateless logger into a wide event Logger class that touches 5 files, split into 3 atomic commits:

Commit	Files	Purpose
1. `refactor: replace logger with Logger class`	`logger.ts`, `types.ts`	Core abstraction + type update
2. `refactor: simplify logging middleware`	`with-logging.ts`	Main consumer of new Logger
3. `feat: enrich wide events with user identity`	`auth.ts`, `index.ts`	Auth integration + re-export

Each commit:

Compiles independently
Contains one logical change
Can be reviewed in isolation
Tells a clear story when read in sequence