← ClaudeAtlas

guardlisted

Run /guard to decide what this agent may run on its own: database writes, migrations, and destructive git. Runs twice in a project's life, once at the start for the rules every project needs, and again once the stack is known to cover that stack's migration tools. Writes the rules into .claude/settings.json so the harness enforces them.
rishabhsisodiya/stop-guessing · ★ 0 · Code & Development · score 70
Install: claude install-skill rishabhsisodiya/stop-guessing
## What this skill does Sets the safety boundary for a project and writes it where the harness enforces it. An agent that can run `prisma migrate reset` or `git reset --hard` will eventually run one at the wrong moment. A skill instruction does not prevent that, because a skill is a prompt and a prompt can be forgotten in a long session. `.claude/settings.json` is not a prompt. ## Two layers, because the stack is not known on day one A greenfield project has no stack until `/architect` decides one. Writing Prisma rules onto a project that turns out to be Django is not protection, it is decoration: it leaves the real tool uncovered while looking like it was covered. That is worse than nothing, because it replaces caution with confidence. So the rules land in two passes: **baseline** — runs first, on any project, no stack knowledge needed. Commands that are wrong everywhere: `sudo`, `rm -rf`, `git push`, history rewriting, `npm publish`, `curl | bash`, production deploys, Docker volume wipes, staging a `.env`, and a lock on the agent editing its own permission files. Plus the read only database gate, which is stack independent by design: it only acts when it actually sees a database client in a command, so it can never create false confidence. **stack** — runs once the stack is known. The migration and schema commands for *that* stack, and nothing else. ### When to run it: once, after the stack is known The recommended flow is **one run, both layers together**, once so