Self-Critique & Revision Loops
Draft, critique against a rubric, revise — and the tricks (fresh context, forced findings) that stop self-review from rubber-stamping.
Get a draft on the table
The loop starts with a normal generation — don't over-engineer this prompt, because the revision passes will do the polishing. Ask for a draft explicitly; framing the output as provisional matters for the critique step.
Draft a cold outreach email to a head of engineering at a 200-person company, introducing our error-monitoring tool. Goal: one reply, not a sale. Under 120 words, no buzzwords, one specific claim about alert noise, one low-friction ask.VerifyYou have a competent draft with visible flaws — probably a generic opener or a soft ask — which is exactly the raw material the loop needs.Critique against a rubric, not vibes
'Any feedback?' invites polite generalities. A rubric with named criteria, forced scoring, and required line citations produces critique you can act on. Forbid rewriting — mixing critique and revision in one step gets you both, done badly.
Critique the email below against each criterion. For each: score 1-5, quote the exact phrase that costs points, and say what would earn a 5. Do not rewrite the email. Criteria: 1. The first line earns the second line (no "I hope this finds you well") 2. Specific and falsifiable beats generic ("cut alert noise 40%" beats "boost productivity") 3. The ask is low-friction and concrete 4. Sounds like a person, not a sequence 5. Under 120 words <email> [paste the draft] </email>VerifyEach criterion gets a score, a quoted culprit phrase, and a concrete fix — no 'overall, solid effort' padding.Revise with the critique as spec
Feed the draft plus the critique back and scope the revision: fix the low scores, preserve the high ones. Unscoped revision requests quietly rewrite everything — including the parts that were working.
Revise the email using the critique. Fix every criterion scoring 3 or below. Do not change what scored 4-5 except where a fix requires it. Keep it under 120 words. <email> [draft] </email> <critique> [critique] </critique> Output only the revised email.VerifyThe weak phrases the critique quoted are gone, and whatever scored well survived intact.Break the rubber stamp
A model reviewing its own fresh output tends to approve it — agreement bias plus in-context anchoring. Three counters: run the critique in a fresh conversation with no authorship trail, assign a hostile-reader persona, and force findings with a quota. 'Name the 3 weakest points' cannot return 'looks good.'
You are a skeptical head of engineering who gets 30 cold emails a week and forwards approximately none. Read this one. Name the 3 things most likely to make you archive it without replying, quoting the exact phrase for each. Then name the single change most likely to earn a reply. Do not be polite. <email> [paste the revised email — in a NEW conversation, without saying you wrote it] </email>VerifyThe fresh-context critic finds real problems the same-thread critique missed — if both passes return applause, your critic setup is too soft.Know when to stop the loop
One or two critique-revise rounds capture most of the available gain; beyond that, prose drifts toward committee-approved mush. And self-critique cannot check claims against the world — factual verification needs sources, tools, or independent checks, not another opinion from the same model.
Loop policy: - Round 1: rubric critique → scoped revision - Round 2: fresh-context hostile critique → scoped revision - Stop when: new critique repeats old points, scores plateau, or edits start swapping good phrasing for different-but-equal phrasing - Never: use self-critique to verify facts, citations, or arithmetic — that needs sources, tools, or independent checks (next lesson: ensembles)VerifyYour second-round diff is visibly smaller than your first — and you can say which remaining issues are factual (needs verification) versus stylistic (the loop's job is done).