fixaiprompt
All templates
⚡ Advanced Techniques·Works on: claude, chatgpt, gemini

Self-Refine — Generate → Critique → Revise (×3)

Technique: Self-Refine

The single move that makes any output 2× better. Used in agentic systems.

Advanced#self-refine#iteration#quality#agents
0/2
Fill the template
2 placeholders left.
Live preview
Act as both author and editor in a Self-Refine loop.

Task: 
Quality bar: 

Run this loop **exactly 3 times**:

**Iteration N:**
1. **Draft N** — write the full output.
2. **Critique N** — switch to a brutal editor's voice. List the 3 most specific weaknesses of Draft N. Not 'could be clearer' — point at the exact sentence and explain what's wrong.
3. **Revision plan N** — for each weakness, the precise change you'll make.
4. Print a one-line scoreline: "Iteration N — drafted ✓, critiqued ✓, ready for revision."

After 3 iterations, print:
- **Final output** (the result of iteration 3's revision plan applied).
- **What changed across iterations** — 2–3 lines comparing v1 to v3.
- **What's still imperfect** — 1 honest line.

Do not abandon the loop early because 'this draft is good enough'. Run all 3.
Inputs0 of 2
e.g. 'sounds like it was written by a person who has done this for 10 years'
›See the lazy version this template replaces
Before — the lazy prompt
Write me <X>.

Why it works

  • Single-shot output uses ~30% of a model's capacity. Self-refine pushes closer to 80%.
  • Forcing a brutal critique voice (different from the writer) breaks the model out of its first-draft attractor.
  • Numbering iterations and printing scoreline prevents the model from cheating by collapsing the loop.
  • Honest 'still imperfect' line at the end keeps the output from over-claiming.

Make this one yours

Replace the bracketed placeholders, then paste into the Prompt Fixer to lint your customisation before hitting send.

Open Prompt FixerBrowse more templates