Summary

Three new git workflow skills adapted from Anthropic's commit-commands plugin. The originals are lightweight commands that run git commit and gh pr create with minimal guidance. These skills take an opinionated approach — convention-aware commit messages, adaptive PR descriptions, user confirmation before destructive actions — and are built as skills rather than commands.

Skills

git-commit — Convention-aware commit creation. Checks loaded project instructions first (AGENTS.md, CLAUDE.md), falls back to recent commit history patterns, defaults to conventional commits. Includes lightweight logical commit splitting at the file level when changes have clearly distinct concerns (no git add -p surgery).

git-commit-push-pr — Produces PR descriptions that scale in depth with the change:

Key behaviors:

• Leads with value ("Fixes timeout during batch exports") not files ("Updated handler.py")
• Design decisions section for medium+ PRs explaining why, not just what
• Conditional test plan — included only when testing approach is non-obvious, omitted for straightforward changes
• No #-prefixed list items (prevents GitHub issue auto-linking)
• No empty template sections

git-clean-gone-branches — Adapted from the original clean_gone command with several fixes: runs git fetch --prune first (original didn't), asks for user confirmation before deleting (original auto-deleted), uses a shell script with safer regex parsing, and handles worktree cleanup.

Adaptations from the originals

• All three use command git / command gh to account for shell aliases and proxies like Rust Token Killer (rtk)
• Built as skills (progressive disclosure, bundled scripts) rather than one-shot commands
• git-commit-push-pr adds the entire adaptive PR description system (the original just called gh pr create with no body guidance)
• git-commit adds convention detection and logical commit splitting (the original had no message guidance)

Eval results

Tested git-commit-push-pr across 8 scenarios using the skill-creator eval framework (with-skill vs without-skill baseline):

• Brevity calibration: small-bugfix description went from ~1000 chars to ~330 chars after iteration
• Design decisions: consistently present in medium+ PRs, absent where unwarranted
• Conditional test plan: correctly omitted for obvious cases, included for complex scenarios (webhook retry with backoff timing, timeout handling, idempotency verification)
• Commit splitting: correctly split mixed concerns, kept single commit for coherent features and ambiguous groupings