claudeframework: Skill for Claude Code

Create one or more commits from the current working tree changes.

$ARGUMENTS is optional. Pass --all to skip the diff analysis and commit everything as a single commit.

Steps

1. Run git status to see staged and unstaged changes
2. Run git diff HEAD (staged + unstaged) to understand the full scope of changes
3. Analyse the diff for distinct logical concerns — ask: would a reviewer understand this as one coherent change, or are there multiple independent things happening?

If multiple distinct concerns are detected (and --all was not passed):

• List the concerns found (e.g. "engine logic change", "new test file", "updated README")
• Propose a split: one atomic commit per concern
• For each proposed commit: describe what it covers and suggest the message
• Ask the user to confirm the split or adjust it
• After confirmation, guide through staging and committing each group in sequence

If a single concern (or --all was passed):

• Proceed to the next step

1. Read project commit conventions: Check if .claude/rules/branch-management.md exists. If it does, read it to understand project-specific branch naming and commit message conventions. Then:

   • Run git branch --show-current to get the current branch name
   • Attempt to parse the branch name according to the documented convention to extract a ticket/issue identifier (e.g. regex ^(\d+)\.(Bug|Improvement|Feature)\.([^.]+)\.(.+)$ extracts ticket ID from segment 1)
   • If the full pattern does not match, try a fallback pattern ^(\d+)\.(.+)$ to extract at least the ticket number
   • If a ticket ID is found, note it for use in the commit message (step 8)
   • If no ticket ID can be parsed and the project rules require one, ask the user for the ticket ID before proceeding
   • If no branch-management.md exists, skip this step

2. Backlog check: if the commit message references a task/PRD identifier (e.g. PRD-17, TASK-5), search for a backlog or tracker file in .claude/input/ that contains that identifier as a heading. If found and the heading does not already have a ✅ marker, add ✅ to the heading and include the file in the commit. Skip this step if no backlog file exists.

3. Run pre-commit checks in this order — read the commands from CLAUDE.md (skip any not listed): a. Typecheck (e.g. tsc --noEmit) b. Lint (e.g. npm run lint)

   • Do NOT run tests or build here — those are for /4_test, not commit gates
   • If a check fails: report the failure and ask the user whether to fix it first or proceed anyway

4. Stage the appropriate files:

   • If specific files were agreed for this commit, stage only those: git add <files>
   • If committing everything: git add -A

5. Compose the commit message using conventional commit format:

   <emoji> <type>(<scope>): <short description>
   

   • Keep the first line under 72 characters
   • Present tense, imperative mood ("add feature" not "added feature")
   • <scope> is optional — use the primary directory or module affected (e.g. scoring, ui, store)

   Type → emoji mapping:

   Type	Emoji	Use when
   feat	✨	New feature or capability
   fix	🐛	Bug fix
   docs	📝	Documentation only
   refactor	♻️	Code restructured, no behaviour change
   test	✅	Tests added or updated
   chore	🔧	Build, config, tooling, dependencies
   perf	⚡️	Performance improvement
   style	🎨	Formatting, whitespace, no logic change
   ci	🚀	CI/CD pipeline changes
   revert	⏪️	Reverts a previous commit

   Project-specific ticket suffix: If a ticket ID was extracted in step 4, append #<ticketId> to the end of the commit message. Ensure the # is NOT at the start of the message (git treats leading # as a comment). The final format becomes:

   <emoji> <type>(<scope>): <short description> #<ticketId>
   

   Example: ✨ feat(ui): add loading spinner to search results #1234