Reflection is self-consistency for prose: instead of sampling N times, sample once, then critique, then revise. The critic pass is where the quality comes from. Frontier models critique their own output harder than their initial draft, catching errors they'd have committed to on a single shot.
Implementation variants: "draft → self-critique → revise" in one call (cheap, works OK), or separate draft / critic / revise agents (more expensive, more reliable). Either way: the prompt to the critic matters. "Look for errors" underperforms "Check each of these 5 criteria; any fail = regression."
Example Prompt
You will write a customer apology email in three steps.
STEP 1 - DRAFT: write the email.
STEP 2 - CRITIQUE: against these criteria, mark each pass/fail:
- Takes ownership without excuse-making
- Offers a specific remedy
- Commits to a follow-up timeline
- Avoids corporate-speak cliches
STEP 3 - REVISE: rewrite to fix any fails from step 2.
Only output the final revised email.When to use it
- Prose output where a second pass meaningfully improves quality
- Tasks with explicit acceptance criteria you can bake into the critic
- Latency can absorb 1-2x the base call cost
When NOT to use it
- Short / rote outputs -- the critique overhead swamps the gain
- No clear evaluation criteria -- the critic hallucinates flaws
- Hard latency SLOs