Skip to content

AI Compass / Knowledge cards

Separate building and checking

a critic with fresh context

Checked Automation Prompting

Whoever built something is poorly placed to see their own mistakes. AI agents are similar: the agent that builds has its entire solution path in its context and considers it correct. A critic is a second agent that starts without that history, that is, with fresh context. It checks the result against a yardstick that was set before the build. Jack Roberts shows the pattern as a “gauntlet loop” for design, AI LABS as check skills and a second opinion for code.

What it is for

Mistakes the builder overlooks get noticed before you accept the result. In the end you have to check less yourself.

Steps

  1. Set the yardstick before the build. For code, that is the original requirements. For design and copy, according to its README the gauntlet-loop repo requires a named model example that the critic can retrieve, such as a particular page or repo. A general grade like “award-worthy design” is not enough.
  2. Set a checking angle for each critic. Jack Roberts names three: Is the brief fulfilled? Are the design rules followed? Is the craftsmanship clean?
  3. Start each critic with fresh context, as a subagent or as its own session. For a second opinion, AI LABS uses claude -p to start a new session without the previous history.
  4. Have it check against the original requirements, not just against one part. An agent that sees only its own part overlooks gaps between the parts.
  5. Have it compare rather than grade. According to the gauntlet-loop README, the critic places the result and the model example side by side without names and picks the better one, because scores rise from round to round.
  6. Build, check, improve, until all critics agree. Set a maximum number of rounds yourself, see Limits. In Jack Roberts’s example it was up to 10 rounds.
  7. For code, use the built-in tools first: according to the docs, Claude Code includes /code-review (finds bugs), /simplify (cleans up) and /security-review (looks for security vulnerabilities). You only need to build what is missing. At AI LABS, that is a check skill for design.

Limits

Sources

Status

Checked 17 October 2026 · Maturity new (pattern available, check commands in the docs; costs and effect only from the creators)

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