| name | writing-commit-messages |
| version | 1.4.0 |
| description | Use when the user asks to generate, write, draft, or validate a commit message, squash commit, or merge commit message. Detects mode (staged, squash, rewrite) from arguments and message context. Do NOT activate during implementation work; only when the user is ready to capture a finished change or validate an existing one. |
| model | sonnet |
| effort | high |
Writing Commit Messages
Generate and validate Conventional Commits with a mode-aware, gather-script-driven workflow.
Output scope: Presents formatted commit message text. Does not create commits, write files, or modify git state.
Workflow
The workflow below can be extended by content earlier in context. Two shapes are recognized:
- Pre-Step / Post-Step instructions. Before executing Step N, check whether earlier context contains a section headed
## Pre-Step-N. If it does, execute its content as additional instructions, then continue with Step N. After Step N, do the same check for ## Post-Step-N.
- Named-value assignments. The skill body cites certain configuration values by backticked name alongside an inline default (for example
`subject.max_length` with default 72). When such a name appears, check whether earlier context assigns a value to it. If yes, use the assigned value; otherwise use the inline default.
Both checks default to no-op. When earlier context contains no matching section or assignment, the skill runs entirely on the defaults documented inline.
digraph writing_commit_messages {
"Start" [shape=doublecircle];
"Detect mode" [shape=box];
"Gather diff via gather.sh" [shape=box];
"Orient: status + shortstat" [shape=box];
"Prioritize files by churn" [shape=box];
"Has log section?" [shape=diamond];
"Read prior context" [shape=box];
"Inspect content per priority" [shape=box];
"Query cross-cutting signals (Grep)" [shape=box];
"Determine type" [shape=box];
"Infer scope" [shape=box];
"Craft subject and body" [shape=box];
"Anti-slop validation" [shape=box];
"Present message with footer" [shape=box];
"Validation mode?" [shape=diamond];
"Run validation checks" [shape=box];
"Offer clipboard copy" [shape=box];
"Cleanup tmpfile" [shape=doublecircle];
"Start" -> "Detect mode" -> "Gather diff via gather.sh" -> "Orient: status + shortstat" -> "Prioritize files by churn" -> "Has log section?";
"Has log section?" -> "Read prior context" [label="yes"];
"Has log section?" -> "Inspect content per priority" [label="no"];
"Read prior context" -> "Inspect content per priority";
"Inspect content per priority" -> "Query cross-cutting signals (Grep)" -> "Determine type" -> "Infer scope" -> "Craft subject and body" -> "Anti-slop validation" -> "Present message with footer" -> "Validation mode?";
"Validation mode?" -> "Run validation checks" [label="yes"];
"Validation mode?" -> "Offer clipboard copy" [label="no"];
"Run validation checks" -> "Offer clipboard copy" -> "Cleanup tmpfile";
}
Step 1: Detect mode
The skill picks one mode in this order. The argument is authoritative: select mode directly from it; never re-interpret it against conversation context (e.g., do not fall back to staged mode because uncommitted changes exist). No slash commands.
- Argument is a full or short SHA → rewrite mode (range =
<sha>^..<sha>).
- Argument is a range expression (
<a>..<b> or <a>...<b>) → squash mode (range as-is).
- User message mentions "rewrite", "reword", or "amend" without an explicit ref → rewrite mode (ref =
HEAD).
- User message mentions "squash", "branch", "PR", or asks for a summary of a branch → squash mode. Base ref is
modes.squash.default_base; default if not otherwise stated: main.
- Otherwise → staged mode (range =
--cached; fall back to working tree git diff if --cached returns empty).
Detection of "validate", "check", "verify", or "is this commit message correct" sets validation mode as an additional flag on top of rewrite or squash mode. Validation mode requires an explicit ref.
Step 2: Gather diff material
Run the gather script by its absolute path (<skill-dir>/scripts/gather.sh), and keep the working directory inside the repository whose changes you are summarizing. Do not cd into the skill's own directory: the script detects its target repository from the current working directory, so running it from anywhere else targets the wrong repository.
bash <skill-dir>/scripts/gather.sh <range>
The script writes all git output to a single file in /tmp and prints a table of contents on stdout.
The range depends on mode:
| Mode | Range |
|---|
| staged | --cached (or empty for working-tree fallback) |
| squash | <base>..HEAD |
| rewrite | <sha>^..<sha> |
Record the TMPFILE path and TOC printed on stdout. Each TOC entry has the form SECTION <name> <start>-<end>. Sections whose range is inverted (start > end) are empty; skip them.
Exit-code handling:
0: success, proceed.1: no changes. For staged mode, retry with empty range. For squash mode, report "branch has no commits ahead of base" and stop. For rewrite mode, report "hash resolves to no changes" and stop.2: invalid range, or not inside a git repository. Report the script's stderr message verbatim to the user and stop.
Step 3: Orient
Read the plugins, status, and shortstat sections together with one Read call. This loads the file change list with rename detection and the one-line summary.
Step 4: Prioritize
Read the numstat section. Rank files by churn (added + deleted lines) and status (Modified > Added > Renamed > Deleted) to identify which files need semantic inspection in Step 6.
Step 5: Read prior context
Run only when the TOC contains a log section (squash and rewrite modes; staged mode skips).
Read the log section. Each commit entry has format:
<sha>
<subject>
<body lines>
---
- Squash mode: per-commit subjects and bodies seed the draft body. Draft the summary from subjects first; later steps verify and fill gaps.
- Rewrite mode: the existing commit message is a starting reference, but base the new type and scope purely on the diff content (Step 6 onward).
Step 6: Inspect content
For each file the draft will make a specific claim about, read the corresponding DIFF_FILE range from the TOC. Use Read with offset = mark_start, limit = mark_end - mark_start + 1.
Every sentence in the subject and body must be traceable to a hunk read here. Content is the source of truth; do not infer behavior from file paths or change counts alone.
Prioritize files in Step 4's order. Stop reading once you have enough evidence for the type, scope, and subject. Typically: top 1-3 highest-churn files for staged/rewrite, top 3-7 for squash.
Step 7: Query signals
Use Grep with path=$TMPFILE for cross-cutting checks that don't belong to one file:
- Removed exported symbols:
^-func , ^-type , ^-class , ^-export , ^-pub fn
- Breaking change markers in text:
BREAKING, breaking change
- Specific identifiers mentioned in the draft (verify they actually appear in the diff)
Grep keeps the diff authoritative while loading only matched lines.
Step 8: Determine type
Apply the decision tree in references/type-detection.md.
Confidence handling:
- HIGH / MEDIUM: use type directly.
- LOW: use
AskUserQuestion with options derived from the analysis.
Detect breaking-change indicators from Step 7 and from API-shape diffs in Step 6.
Step 9: Infer scope
Determine scope from the file list (from Step 3's status section). Universal default: the immediate top-level directory of changed files is the scope. When changes span multiple directories, choose the one with the most files; ties prefer the directory with the most line churn. Omit scope when changes are repository-wide or when no single directory dominates.
Check the resolved scope against scope.naming_convention. Default if not otherwise stated: kebab-case (matches [a-z0-9-]+).
Confidence handling:
- HIGH / MEDIUM: use scope directly.
- LOW: use
AskUserQuestion.
Step 10: Craft subject and body
Subject rules:
- Imperative mood, present tense ("add", not "added").
- Lowercase first word after the colon.
- No period at end.
- Hard cap:
subject.max_length characters including the type(scope): prefix. Default if not otherwise stated: 72.
- Specific: name the component and behavior, not abstract descriptions.
- No PR number in subject.
- No AI/tool attribution.
Body rules (when a body is included):
- Blank line between subject and body.
- Single continuous lines per paragraph (no hard-wrapping at 72).
- WHY-not-WHAT: motivation, root cause, trade-offs.
- Concrete: class names, config keys, method signatures.
- For breaking changes: include a
BREAKING CHANGE: footer. When a breaking-change footer is emitted, append body.breaking_change_handoff to the body. Default if not otherwise stated: include migration instructions inline.
Squash body: describe the branch's purpose grouped by logical concern, not by commit order. Each Step 5 commit contributes a candidate concern; merge overlapping ones.
Rewrite body: if the existing message had a body and Step 6 inspection confirms its claims, preserve substance; otherwise rewrite.
Step 11: Anti-slop validation
- Check each body paragraph for hard-wrapping. Join multi-line paragraphs into a single continuous line. The 72-char limit applies only to the subject.
- Dispatch the
human-author:ai-slop-writing-fixer subagent via the Agent tool. Pass the full message text (subject and body, including any breaking-change footer but excluding the Co-Authored-By and other generated footer lines from Step 12) as content. Use the returned fixed_content as the message text for Step 12. If the agent's changes list is non-empty, the violations are already corrected; do not re-apply them.
Step 12: Present message with footer
Quick self-check: type matches changes, scope matches files, subject is accurate.
Output:
- Brief analysis (type reasoning, scope reasoning, breaking-change note if any).
- The commit message in a fenced code block.
Message format:
type(scope): subject
body (when a body is included)
BREAKING CHANGE: description (when breaking)
<footer lines>
Footer: emit footer.template with {model} substituted by the active model's name (e.g., Opus 4.7 (1M context), Sonnet 4.6). Default if not otherwise stated: Co-Authored-By: Claude {model} <noreply@anthropic.com>. If footer.template is the empty string, suppress the template line.
After the template line, emit each entry of footer.extra_lines on its own line. Default if not otherwise stated: no extra lines.
Do NOT include PR references like (#N) in the subject. GitHub adds these during merge.
Step V: Validation checks
Run only in validation mode. Reuse Steps 2-7 (gather, orient, inspect, grep) to build the "what the commit should say" baseline.
Apply checks per references/validation-rules.md.
Report verdict per category, then overall. Optionally generate a rewrite suggestion (Steps 8-12 against the actual diff) and present side-by-side.
Step 13: Offer clipboard copy
Ask the user whether to copy the generated message to the clipboard. If they accept, use the clipboard_copy MCP tool (or pbcopy / xclip fallback if MCP is unavailable). Copy only the message content; do not include the surrounding code fence markers.
In validation mode, the offer applies to the rewrite suggestion (if generated), not the report.
Step 14: Cleanup
Delete the tmpfile as the final action, regardless of outcome, using the cleanup script by its absolute path:
bash <skill-dir>/scripts/cleanup.sh "$TMPFILE"
This step runs on every termination path, including error exits from earlier steps.
