git-master

SkillDev tools

Use when committing changes to git or naming branches. Triggers include "commit", "커밋", "git commit", "finalize changes", "save to git", "commit my work", "branch name", "브랜치 이름", "what should I name this branch".

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the git-master skill

What this skill tells your AI

The instructions your AI receives, as published by toongri/oh-my-toong-playground in skills/git-master/SKILL.md and read by ahel’s review.

Git Master

Analyze code changes and generate Korean commit messages following project conventions. Also applies branch naming conventions when naming new branches.

"A good commit makes history easy to read. A bad commit turns git log into a graveyard."


<Critical_Constraints>

The Iron Law

NO COMMIT WITHOUT:
1. Single logical change (or properly split)
2. Message ≤ 50 characters
3. Subject comprehensible to git log readers without external context

Violating the letter of these rules IS violating the spirit.

There are no exceptions. User preference does not override project rules.


Non-Negotiable Rules

These are RULES, not guidelines. This project enforces them strictly.

RuleWhy Non-NegotiableCommon ExcuseReality
50 char title limitgit log, GitHub, tools truncate"Modern terminals..."Tools haven't changed. 50 chars.
One logical changeAtomic = reviewable, revertable"Save my work"Use branches, not mega-commits
Korean 명사형 종결Project convention"I prefer English"Project rules override preference

</Critical_Constraints>


Workflow Files: NEVER COMMIT

These files are workflow artifacts, not implementation deliverables:

# ALWAYS unstage these before committing:
git reset HEAD plan.md 2>/dev/null || true
git reset HEAD research.md 2>/dev/null || true
git reset HEAD docs/specs/ 2>/dev/null || true

Why?

  • plan.md: Worker updates separately after commit
  • research.md: Research artifact, not code
  • docs/specs/*: Input documents, never modified

If user says "I always commit plan.md with my code" → Refuse. Project rules.


Best Practice Additions (Industry Standard)

Imperative Mood

Commit messages should be written in imperative mood. The message should complete the sentence: "If applied, this commit will [your message]".

GoodBad
feat: 사용자 인증 기능 추가feat: 사용자 인증 기능 추가함
fix: 결제 오류 수정fix: 결제 오류 수정했음

Breaking Changes

Changes that break backward compatibility must be marked:

MethodFormatExample
Type with exclamationtype + "!" + messagefeat!: API 응답 형식 변경
FooterBREAKING CHANGE: descWrite at the end of body

Git Trailers (Optional)

Add trailers at the end of body when needed:

TrailerUsage
Co-authored-by:Pair programming
Fixes:Issue linking (Fixes: #123)
Signed-off-by:Projects requiring DCO

Core Principle

One commit = One logical change

  • Separate unrelated changes (atomic commits)
  • Title must be within 50 characters, core message only
  • WHY goes in body (optional)
  • Many files ≠ many commits — logical cohesion decides (see Atomic Commit Splitting)

Product, Not Process — Commit messages describe the change itself.

  • Describe what changed, rather than the source of the change (code review, issue number, meeting decision)
  • Someone reading git log six months later will not know what "리뷰 P1-3 수정" means
  • Record the source/context in the body or a trailer (Fixes #123)

Quick Reference

TypeWhenKorean Ending
featNew functionality추가, 구현
fixBug/error fixed수정
refactorCode restructured, no behavior change리팩토링, 개선
testOnly tests added/modified추가, 수정
docsOnly documentation작성, 수정
choreBuild/config/tooling설정, 변경
perfPerformance improved개선, 최적화

Process Steps

Step 1: Analyze Changes

git status
git diff
git diff --staged

For each changed file, categorize:

  • What domain/feature is affected?
  • What is the main change?
  • Are there multiple logical changes? → Split!

Atomic Commit Splitting

"Separate each logical change into a separate patch." — Linux Kernel Documentation

"If you make a single change to numerous files, group those changes into a single patch." — Linux Kernel Patch Philosophy

File count is a TRIGGER for analysis, NOT a splitting rule.

Split Analysis Trigger
Changed FilesAction
1-2 filesLikely single commit — verify one logical change
3+ filesPause and analyze — are there multiple concerns?
10+ filesStrongly consider splitting — multiple concerns are probable
Mandatory Self-Check (3+ Files)

When 3+ files change, you must perform this self-check before committing:

"N개 파일을 M개 커밋으로 만든다."
IF M == 1 AND N >= 3:
  → Is this really one logical change?
  → Can you explain in one sentence why these files must stay together?
  → If you cannot → SPLIT

This is not a numerical formula. 3+ files means "think it through," not "always split." A single commit is justified when logical cohesion is sufficient, as in Example 7 (point accrual: 4 files = 1 commit).

Commit Justification (3+ Files per Commit)

When a commit includes 3 or more files, you must explain why they belong together in one sentence:

"이 커밋은 [파일들]을 포함한다. 이유: [구체적 이유]"
Valid ReasonInvalid Reason (→ must split)
Implementation + its direct test file"같은 기능 관련" (vague)
Type definition + its only consumer"같은 PR에 포함" (not a reason)
Migration + model change (splitting breaks the build)"함께 변경됨" (not a reason)
Multiple files in a single rename operation"관련 있어서" (vague)

IMPORTANT: One feature ≠ one commit. A feature may contain multiple logical changes (config, domain, service, test, docs). Each independently meaningful layer is a separate commit. However, a single atomic operation (e.g., renaming across 10 files) IS one commit.

When to Split

Split when ANY of these are true:

SignalExample
Different change types mixedBug fix + unrelated refactor
Different domains/modules affectedauth/ change + user/ change with no dependency
Independently revertable partsConfig change that works without the feature using it
Description gets too long"Fixed X and also added Y and refactored Z"
Different architectural layersConfig + domain + service + test + docs for one feature
Multiple independent changes (even in 1-2 files)3 review findings each require an independent change → 3 commits
When NOT to Split

Keep as single commit when:

SignalExample
Truly atomic operationRenaming a class across 5 files
Tightly coupled pairDTO definition + the single mapper using it
Cannot exist independentlyInterface + its only implementation (in same module)
Single mechanical changeFormatting/linting across many files
Grouping Strategy (when splitting)

Commit in this order (dependency-first):

  1. Config/Build — dependencies, build settings
  2. Infrastructure — refactoring, API changes
  3. Source/Logic — business logic, features
  4. Tests — related test code
  5. Documentation — README, docs
Splitting Rules

Each split commit must:

  • Be independently meaningful (not "part 1 of 3")
  • Have its own proper commit message
  • Leave the codebase in a buildable state
  • Be revertable without breaking other commits
Test-Implementation Pairing

Test files must be included in the same commit as their corresponding implementation:

Test PatternImplementation File
*_test.sh*.sh
*.test.ts*.ts
*.spec.ts*.ts
*Test.kt*.kt
__tests__/*Corresponding source

Anti-pattern: Splitting implementation and tests into separate commits. An implementation commit without tests cannot be verified in its intermediate state.

Step 2: Verify No Workflow Files

git diff --staged --name-only | grep -E "^(plan\.md|research\.md|docs/specs/)"

If any match → Unstage them before proceeding.

Step 3: Determine Commit Type

  • New functionality → feat
  • Bug/error fixed → fix
  • Code restructured without behavior change → refactor
  • Only tests → test
  • Only docs → docs
    • Functionality vs documentation: "If it defines system behavior, it is functionality; if it provides reference/shared information for human readers, it is documentation"
    • Changes to functional files (SKILL.md, agents/.md, rules/.md, hooks/*) → feat/fix/refactor
    • Changes to documentation files (README.md, API specifications, guides) → docs
  • Build/config → chore
  • Performance → perf

Step 4: Output Commit Plan (3+ Files — BLOCKING)

When 3 or more files change, you must output a commit plan before executing any commit:

COMMIT PLAN
───────────
변경 파일: N개
계획 커밋 수: M개

COMMIT 1: type: 제목
  - path/to/file1
  - path/to/file1_test
  Justification: 구현체 + 직접 테스트

COMMIT 2: type: 제목
  - path/to/file2
  Justification: 독립적 설정 변경

실행 순서: Commit 1 → Commit 2
(의존성 순서: Config → Source → Test → Docs)

Do not proceed to commit execution without this output. Skip this step for changes to 1-2 files.

Caution about fix: Changes arising from code review are not all fix. Even for review findings, use feat for new functionality and refactor for structural improvements. Only actual bug/error fixes use fix.

Step 5: Generate Commit Message

MANDATORY Self-Check (Immediately After Drafting the Subject)

Immediately after drafting the subject and before committing, you must check it for invented/opaque labels. If any pattern matches, rewrite and check again. This check enforces anti-pattern 1 (Invented/opaque label ban) of the communication-style rule (rules/communication-style.md). The canonical source of the detection regexes is hooks/lib/label-patterns.sh — git-master does not keep its own copy.

Why this is mandatory: Plan step numbers and AC IDs are clear to the worker viewing the plan, but git log readers cannot access that plan — past violations have required history rewrites to correct them.

When a violating pattern is found:

  • Simply remove the token: (Step 12) → delete (when the subject is already self-contained in domain terms)
  • Replace the token with domain terms: align RN tooling lockstep with mobile (Step 4)RN tooling mobile에 정렬
  • If traceability is truly needed, move it to a body trailer: Refs: dispenser-monorepo-absorption.md#step-12

Subject rules (NON-NEGOTIABLE):

  • Korean (한국어)
  • Max 50 characters ← ENFORCED, not a guideline
  • 명사형 종결 (e.g., "추가", "수정", "삭제", "구현", "개선")
  • No period at end

Subject content rule:

The subject's audience is a future git log reader — someone encountering the line six months later or another developer doing code archaeology. The subject must let that reader understand what changed without external context.

Reader model:

What Readers HaveWhat Readers Do Not Have
The codebase itselfPR description, review thread
commit body / diffWork session context
Domain knowledgeInternal classification systems (P-ratings, severity labels)
History of other commitsMeeting notes, Slack messages

Validation questions — ask yourself after writing the subject:

  1. "Can readers understand what changed from this subject alone?"
  2. "Do readers need access to external documents/session context to understand it?"

If 1 is NO or 2 is YES → rewrite.

Common failure patterns (depend on external context):

PatternWhy It Fails
Review classifications (P0/P1/HIGH/CRITICAL, etc.)Readers cannot access the classification definitions
Workflow labels (잔여/residual/follow-up)Session context is needed to know what remains
Process references (리뷰/audit/라운드)External documents are needed to identify the review/audit
Vague counts (3건/여러 건 alone)What the count refers to is unclear without the body
Plan step numbers (Step N/Step 7.6/Phase N/Round N)The external plan is needed to identify which plan's step N
Acceptance criteria IDs (AC M1/H4/(M3))AC definitions are in an external plan/spec inaccessible to readers

These are clear to you during the work but meaningless to git log readers. Put any needed source/classification/count in the body or a trailer — describe the change itself in domain terms in the subject.

BAD vs GOOD subjects (real examples):

BAD (depends on external context)GOOD (self-contained, domain terms)
fix: collect-jd P1 스펙 드리프트 3건 정합fix: ledger filename + canonical path + Gate 5 classification 정합
refactor: SKILL.md HIGH 잔여 3섹션 cross-ref 전환refactor: SKILL.md Session Lock + Atomic Write + L1/L2 cross-ref 전환
fix: 코드 리뷰 P1/P2 이슈 수정fix: persistence 저장 시점을 Step 완료 단위로 변경
chore(dispenser): remove per-app husky (AC M1)chore(dispenser): per-app husky 제거

GOOD subjects directly reveal the changed area (file/module/domain concept) without external documents.

If subject > 50 chars:

  1. Identify the ONE core change
  2. Remove unnecessary words
  3. Move details to body
  4. Do NOT commit with > 50 chars

Body rules (when needed):

  • Blank line between subject and body
  • Wrap at 72 characters
  • Explain WHY, not WHAT

Footer rules (when needed):

  • Blank line between body and footer
  • BREAKING CHANGE: description for breaking changes
  • Fixes #123 for issue references
  • Co-authored-by: Name <email> for pair programming

See references/commit-conventions.md for complete format.

Step 6: Execute Commit

git add .
git reset HEAD plan.md 2>/dev/null || true
git reset HEAD research.md 2>/dev/null || true
git reset HEAD docs/specs/ 2>/dev/null || true

# Verify staged files
git diff --staged --name-only

# Commit
git commit -m "$(cat <<'EOF'
type: 한국어 제목 50자 이내
EOF
)"

Step 7: Return Result

## Commit Result
- **Hash**: [7-char hash]
- **Type**: [feat/fix/refactor/etc.]
- **Message**: [full commit message]
- **Files**: [count] files changed

Branch Naming Convention

Format: <type>/<description> (kebab-case, English)

TypeWhen to Use
feature/New functionality
fix/Bug fixes
refactor/Code restructuring
chore/Build, config, tooling
docs/Documentation only
test/Tests only

Examples: feature/user-auth, fix/login-redirect, refactor/api-middleware, chore/update-deps

Rules:

  • All lowercase, words separated by hyphens
  • Description is English, concise (2-4 words)
  • No special characters except hyphens

Edge Cases

No changes: Return "Warning: No changes to commit. Working tree is clean."

Message too long: NEVER just "accept" a long message. Shorten it.

Mixed types: Use primary type, mention secondary in body.

User insists on violation: Explain why you cannot comply. Offer alternatives.

Large cohesive change (10+ files): Analyze by concern. One feature ≠ one commit. Split by architectural layer (config, source, test, docs) unless the change is a single atomic operation (e.g., rename).


Examples

See examples.md for commit message examples.


Common Mistakes

MistakeWhy It's WrongFix
Multiple features in one commitHard to rollback/cherry-pickSeparate by logical unit
Vague messages like "수정함"Unclear what was changed and whyDescribe specific changes
Writing commit messages in EnglishProject convention violationUse Korean 명사형 종결
Period at end of titleUnnecessary characterRemove period
Title exceeding 50 charactersTruncated in git logKeep core message, move details to body
Committing plan.mdWorkflow files mixed ingit reset HEAD plan.md
Meta-commit: "리뷰 이슈 수정"The change is opaque; git log becomes meaninglessDescribe the actual change: "저장 시점을 Step 완료 단위로 변경"
Opaque reference: "P1-1, P2-3 반영"Cannot be decoded without external documentsPut references in the body/trailer and the change itself in the subject
Subjects depending on external context (P1 X, HIGH 잔여 Y, 리뷰 N건)git log readers cannot access the classification system/session context — meaning is lostDescribe the change itself in domain terms; put classification/context in the body or trailer
Embedding plan-step / AC IDsgit log readers cannot decode them without the plan — tokens meaningful only to the workerAutomatically check with Step 5 MANDATORY Self-Check (canonical patterns: hooks/lib/label-patterns.sh); put traceability in the PR description or trailer

When NOT to Use

  • Uncommitted changes in unrelated files → Stash or separate commit
  • No actual changes → Nothing to commit

Signals

GitHub stars
25
Forks
1
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
git-master-toongri
Source
github.com/toongri/oh-my-toong-playground