git-workflow-master

Trigger phrases

Phrases that activate this skill when typed to Claude Code:

• clean up this branch
• rebase before merge
• set up a git workflow

What it does

git-workflow-master is the version-control persona in the agency-agents collection. It picks the right branching strategy for the team (trunk-based by default; Git Flow only when versioned releases require it), enforces atomic conventional-commit history (feat:, fix:, chore:, docs:, refactor:, test:), and routes every potentially destructive operation through its safe variant — --force-with-lease instead of --force, interactive rebase before merge, and meaningful branch names like feat/user-auth rather than mybranch2.

Five rules govern the persona’s output: atomic commits that revert independently, conventional format, never force-push shared branches, always rebase on the target before merging, and meaningful branch names. Recovery steps ship alongside risky operations — the persona explains the reflog before suggesting anything that touches it.

When to use it

• Cleaning up a branch with WIP commits before opening a PR (squash, reword, drop)
• Resolving a merge conflict where you want the rebase-vs-merge tradeoff explained
• Standing up a branching strategy for a new repo or new team
• Recovering a “lost” commit (detached HEAD, force-pushed branch) where reflog or git fsck is the answer
• Pairing with using-git-worktrees from obra/superpowers — the worktree skill creates isolation, this one keeps the history clean inside it

When not to reach for it:

• One-line fixes on main where an atomic commit is already what you’d write
• Repos where the team has explicitly chosen non-conventional commit messages (don’t fight house style)
• Projects using Gerrit, Jujutsu, or other non-Git VCS — the persona’s mental model assumes Git primitives

Install

From msitarzewski/agency-agents at engineering/engineering-git-workflow-master.md. Copy to ~/.claude/agents/ or use the repo’s installer. Standalone — pairs naturally with worktree and code-review personas but doesn’t require them.

What a session looks like

1. Diagnose the state. git status, git log --oneline origin/main..HEAD, git diff origin/main — what’s the actual delta, what does the history look like, is the branch ahead of origin/main cleanly or behind.
2. Pick the operation. Cleanup before PR? Interactive rebase. Resolve conflict? Rebase or merge depending on whether the branch is shared. Recover work? Reflog. The persona names the operation before running it.
3. Show the safe form. git push --force-with-lease instead of git push --force. git rebase -i origin/main instead of git rebase main from a stale base. git branch -d instead of git branch -D.
4. Run it, narrate the diff. Each commit gets a one-line summary; conflict resolution gets file-by-file reasoning rather than blanket “accept theirs”.
5. Recovery on standby. Before any rewrite, the persona prints the pre-rewrite HEAD SHA so reflog recovery is one command away.

The discipline that makes it work: rule 3, “never force-push shared branches.” The persona will refuse a plain --force against an upstream branch and substitute --force-with-lease automatically — the difference between “I overwrote my own work” and “I overwrote someone else’s work.”

Receipts

2026-05-05 — Read-only assessment of the current branch state

Pointed the persona at this repo with two modified files in the working tree (src/layouts/Base.astro, src/content/blog/skill-security-auditor-not-a-website-tool.md) and an explicit “don’t run any destructive operation, just recommend.” Persona ran 5 read-only commands (git status, git diff --stat, git log -10 --oneline, git branch --show-current, git remote -v) and produced a structured assessment.

The receipt-worthy finding was the coherence call: those 2 files are one logical commit, not two. Both come from the same May 4–5 metaDesc smart-truncate session described in session.md (the helper in Base.astro caps SERP descriptions at 160 chars; the blog post lede was padded 72→133 words in the same pass for the AI-citation band). Suggested message: Cap meta descriptions at 160 chars + pad skill-security-auditor lede. Persona then proactively applied the project’s “ask before src/ edits” policy — recommended Hold pending operator go-ahead rather than committing on its own — without being given the policy file by name, just the repo’s existence.

Honest scope: persona was emulated (general-purpose agent + persona’s described framing). The --force-with-lease substitution rule didn’t apply (no force-push needed); the conventional-commit rule and atomic-commit rule both did, and the “name the operation before running it” discipline produced the cleaner-than-default output.

Source and attribution

From Michael Sitarzewski’s agency-agents repository, an MIT-licensed collection of 144+ AI agent personas across engineering, marketing, design, testing, and specialized roles.

License: MIT.

Quote from the persona body, verbatim: “Clean history, atomic commits, and branches that tell a story.” The five rules and the --force-with-lease substitution are how the persona makes that stance operational instead of aspirational.