ade35dc0d8
The repeated "The load-bearing constraint:" opener on every page read as an agent tell. Strip the label across all skill pages and let the constraint stand as a plain declarative sentence; update the writing-docs template so it isn't regenerated. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2.6 KiB
Quickstart:
npx skills add mattpocock/skills --skill=git-guardrails-claude-code
npx skills update git-guardrails-claude-code
What it does
git-guardrails-claude-code installs a Claude Code PreToolUse hook that blocks destructive git commands — git push, reset --hard, clean -f, branch -D, checkout . / restore . — before they run.
This is a guardrail, not a request. The block lives in the harness, not in the agent's instructions — the hook inspects every Bash command and, on a match, exits with code 2 so the command never executes. An agent can forget a "please don't push" note in its prompt; it cannot talk its way past a hook that fires before the tool call lands.
When to reach for it
Type /git-guardrails-claude-code, or the agent reaches for it automatically when a task fits.
Reach for it when you want an agent to work freely in a repo but never perform the handful of git operations that lose work or rewrite shared history. It's a one-time install: run it once per project (or once globally) and the guardrail is standing thereafter. For enforcing quality at commit time — Prettier, type-checking, tests via Husky — that's a different mechanism, so use setup-pre-commit instead.
The guardrail
The hook is a small shell script matching each Bash command against a list of dangerous patterns. When one matches, the agent sees a BLOCKED message telling it that it does not have authority to run that command, and the command is dropped.
Two knobs at install time: scope — this project (.claude/settings.json) or all projects (~/.claude/settings.json) — and the pattern list, which you can extend or trim so the wall sits exactly where you want it. Everything not on the list still runs untouched.
It's working if
- A blocked command (e.g.
git push) never executes, and the agent reports it was denied rather than that it failed. - Ordinary git —
add,commit,status,checkout <branch>— runs as normal.
Where it fits
This is a run-once setup skill, and a standalone: install the guardrail and forget it. Its closest neighbour is setup-pre-commit, which also writes hooks into a repo but guards commit quality rather than destructive git actions — the two stack cleanly. When you're unsure which safety or setup skill fits, ask-matt routes you.