Skip to content

AI Compass / Quick reads

Whoever builds does not check

Automation Prompting

Four eyes help with AI too. The second pair only has to be fresh and know beforehand what it is checking against.

In depth: Separate building and checking – a critic with fresh context

An agent that checks its own work considers its solution path correct, because it is in its context. Jack Roberts therefore separates builder and critic: three critics with fresh context check whether the brief is fulfilled, whether the design rules are right and whether the craftsmanship is clean. The builder improves until all agree, in the example up to 10 rounds (10 August). For code, AI LABS adds: check against the original requirements, not against the part, and get a second opinion via claude -p from a session without history (29 July). Open question is the critic’s model: Jack uses small models, AI LABS warns about false errors from weak checkers.

Tool and alternative

According to the docs, Claude Code includes /code-review, /simplify and /security-review (docs checked). For design, the gauntlet-loop repo provides a ready-made skill file (CC BY 4.0, checked externally). According to the README, the critic compares blind against a model example instead of giving grades. Jack’s own skill is in his community; you do not need it.

Sources

Advertising in the source: Jack Roberts with self-promotion and a Glaido link. AI LABS sponsored by SerpApi, plus a paid community.

Want to try this with your team? That’s what the AI Lab is for. See the AI Lab