The prompt pattern I use most: "critic, then rewrite"

After a couple of years of experimenting with fancy prompt engineering tricks (chain-of-thought, few-shot, persona stacking, XML tags, you name it), the one pattern I reach for more than any other is embarrassingly simple: ask the model to critique its first draft, then rewrite.

Two-step version:
Step 1: "Draft [thing]."
Step 2: "Now list the weaknesses, gaps, and places this could be tighter. Be harsh and specific. Then produce a final version that fixes them."

That's it. It measurably improves almost every kind of output — emails, code reviews, blog posts, summaries, proposals, cover letters, even code. The model is much better at *criticizing* a draft than at producing a perfect one cold, because criticizing doesn't require committing to a structure.

A few variants I use:
- "List every assumption this draft is making. Then rewrite with assumptions stated."
- "What would the strongest possible objection to this be? Rewrite to pre-empt it."
- "Score this draft from 1-10 on [specific criteria, e.g. clarity, actionability]. Below 8, rewrite."
- "Give me 3 versions: A (conservative), B (bold), C (weird). Then pick the best one and tell me why."

I resisted this for a while because it feels like wasting tokens on a first draft, but the delta in quality is large enough that I've stopped optimizing for token count on anything that matters.

What's the deceptively-simple prompt pattern you find yourself reusing? Curious to collect more.

Critic-then-rewrite is my default for everything. Adding to that: having it list the 3-5 core assumptions it's making before answering, so I can correct bad premises. Saved me from garbage outputs dozens of times, especially when I'm asking about something I don't know much about.

The other pattern I use a lot is 'steel man then critique' — makes the output way less one-sided.

---

Critic-then-rewrite is my default for everything. Adding to that: having it list the 3-5 core assumptions it's making before answering, so I can correct bad premises. Saved me from garbage outputs dozens of times, especially when I'm asking about something I don't know much about.

---

Critic-then-rewrite is my default for everything. Adding to that: having it list the 3-5 core assumptions it's making before answering, so I can correct bad premises. Saved me from garbage outputs dozens of times, especially when I'm asking about something I don't know much about (which is where the bad assumptions hide).