The review loop | Master AI Automation in 4 hours Master AI Automation in 4 hours Course About Ayush Modules Sample chapter Toolbox The Microcap Minute Classroom / Module 07: The AI Relay Team / Chapter 5 The review loop Watch first, then read. Same lesson, your pace. What you will learn – Three reviewer personas worth deploying – The loop protocol: iterate until gate passes or budget dies – Self-audit prompts that catch most slop early Review is a role, not a phase Chapter 4 used one reviewer. Real workflows rotate several personas , the same model can wear each hat, though different models add independence (Chapter 1’s reason #2): The Critic , quality gate. “List every sentence that is vague, unsupported, or padded. Be harsh; niceness is failure.” Catches weak writing before humans see it. The Red Team , adversary. “Argue why this plan fails. Attack assumptions; find the cheapest way this embarrasses us.” Surfaces risks authors can’t see. Uncomfortable by design. The Auditor , compliance check. “Verify: every claim has a source or [CHECK] flag; word counts hold; format contract met. Return PASS/FAIL with line numbers.” Mechanical, exhaustive, boring, perfect for machines. The loop protocol Review without rules becomes infinite nitpicking. Bound it: budget = 2 iterations (decide BEFORE starting) for round in 1..budget: issues = reviewer(artifact) # numbered, severity-tagged if verdict == PASS: break artifact = worker(artifact + issues) # fix leg via Handover Doc ship best version you have Two iterations catches most fixable problems; beyond that you’re polishing mud, restart the draft instead (Module 2 Chapter 4’s desk-hygiene rule applies to teams too). And always ship: a reviewed B+ beats an unreviewed A− sitting in drafts forever. Self-audits: cheap first-pass gates Before spending a reviewer leg, make the author audit itself: “Re-read your output against every constraint in section 2. List violations.” “Which claim here are you least sure of? Flag it honestly.” “Cut 20% of the words without losing meaning.” Self-caught errors are free; reviewer legs cost quota. Gate in order: self-audit → critic → red team (if stakes high) → ship. Try it yourself Take Chapter 4’s final newsletter and run all three personas as separate handover legs (mix models freely). Merge findings into one numbered list, run ONE fix leg, re-audit. Record verdicts per round in learn/review-loop-log.md . Round-2 issues should shrink sharply, that’s the gate working. Key takeaways – Personas: Critic (quality), Red Team (risk), Auditor (compliance). – Budget two iterations upfront; always ship the best current version. – Order cheap-to-expensive: self-audit → cross-model review → red team. – Numbered severity-tagged issues make fix legs surgical. Download the exercise sheet (PDF) Module workbook (PDF) ← Prev: Full worked example Next: When teamwork is overkill → Classroom / Module 07: The AI Relay Team / Chapter 5 The review loop What you will learn – Three reviewer personas worth deploying – The loop protocol: iterate until gate passes or budget dies – Self-audit prompts that catch most slop early Review is a role, not a phase Chapter 4 used one reviewer. Real workflows rotate several personas , the same model can wear each hat, though different models add independence (Chapter 1’s reason #2): The Critic , quality gate. “List every sentence that is vague, unsupported, or padded. Be harsh; niceness is failure.” Catches weak writing before humans see it. The Red Team , adversary. “Argue why this plan fails. Attack assumptions; find the cheapest way this embarrasses us.” Surfaces risks authors can’t see. Uncomfortable by design. The Auditor , compliance check. “Verify: every claim has a source or [CHECK] flag; word counts hold; format contract met. Return PASS/FAIL with line numbers.” Mechanical, exhaustive, boring, perfect for machines. The loop protocol Review without rules becomes infinite nitpicking. Bound it: budget = 2 iterations (decide BEFORE startin
The review loop
Written by
in