What problem does it solve? Standards hooks like TDD enforcement and code-rules gates slow down exploratory proof-of-concept builds, but skipping them entirely risks shipping unreviewed code. This Skill separates the two concerns: it runs a fast, hookless sandbox build in an isolated git worktree, then re-verifies the result under full standards before it becomes a commit and pull request. ## Core Features & Use Cases - Hookless sandbox builds: Launches a headless claude -p --bare session in an isolated worktree off origin/main, keeping only two safety gates (personal-data blocking and destructive-command blocking) active. - Verified containment: Builds a minimal safety settings file from the live ~/.claude/settings.json and probes both safety hooks with payloads they must hard-deny before trusting the sandbox. - Clean-room promotion: Re-applies the POC as an uncommitted diff on a fresh branch, runs cleanup, privacy sweep, review and verification, then commits and opens a draft PR with honest limitations stated. - Use Case: You want to spike a caching layer without TDD friction. The sandbox agent builds it freely; if it works, promotion re-verifies the diff, sweeps for secrets, and ships it as a reviewed draft PR. ## Quick Start Ask the AI to prototype your idea, for example: prototype a proof of concept for the widget cache and promote it if it works.
