lbtc

implement-with-review

Implement a requirement, then automatically enter a build-review loop that runs until no genuine bug is left. This instance (Claude) writes the code, judges independently whether each finding raised by the reviewer is real, and fixes only the real ones; the reviewer is whatever the user configured (Codex / Gemini / Grok Build / a separate Claude, or this same claude on a different API key) and it only reports — it never fixes. Use this skill whenever the user asks to "implement / build / write" a feature or requirement and wants cross-review, a code gate, or quality assurance, even if they never say the word "review". Every invocation is a brand-new cycle, fully isolated from past requirements: the reviewer sees only this round's changes and never re-reviews old work.