Make it a critic

The single most common reason a piece of AI output isn't good enough is that the model wrote it in one pass and never looked at it again. The fix is almost embarrassingly simple: in the very next message, ask the model to critique what it just produced. Then ask it to rewrite.

Models are dramatically better at finding problems in a piece of work than at avoiding those problems on a first draft. Editor mode > author mode. Knowing this is the difference between AI work that ships and AI work that gets quietly redone by a human at 11pm.

The basic two-pass

After the model produces a draft, send this:

Critique the draft above. List the three biggest problems with it —
be specific, name passages. Then rewrite the draft addressing those
problems.

That's it. The first pass produces a draft. The second pass — same conversation, same context — produces something noticeably better. You'll see things in the critique you wouldn't have caught yourself: a number that doesn't match the source, a sentence that buries the headline, a tone that's wrong for the audience.

When the work matters, do not ship a one-pass output. Always ask for the critique.

The four critique modes

Different prompts surface different problems. Pick the one that fits the work.

1. The editor pass

For most knowledge-work outputs (reports, summaries, emails, memos).

"Critique this draft. List the three biggest problems with it. Then rewrite addressing them."

Catches: structural problems, missing context, weak openings, redundancy.

2. The red-team pass

For recommendations, hypotheses, and any output that takes a position.

"Argue the strongest case against the recommendation above. Where is the analysis weakest? Which assumption, if wrong, breaks the whole thing?"

Catches: motivated reasoning, single-source dependencies, claims that won't survive a skeptical room.

3. The pre-mortem

For deliverables that will land in front of an audience.

"Imagine this report bombed at the board meeting. List five specific reasons it could have. Be uncharitable."

Catches: audience mismatches, the thing you forgot to address, the question you didn't see coming.

4. The audience-fail check

For anything going to a specific named audience.

"What would a skeptical CFO push back on first? Which sentence would they stop on and ask, 'how do we know that?'"

Catches: unsubstantiated claims, vague verbs, places where you reasoned but didn't show your work.

The chain that ships

For high-stakes outputs, run all three steps:

1. Draft — your normal prompt, normal output.
2. Critique — pick the mode that fits.
3. Rewrite — "Now rewrite the draft addressing the critique. Keep what's working, fix what's broken."

Three messages. Maybe 90 seconds of waiting. The output that lands in your inbox at the end is usually two or three quality grades above the one-pass version.

Why this works

In one-pass mode, the model is doing two jobs at once: producing prose AND remembering not to mess up. It's distracted. Quality drops on both.

In editor mode, the model is doing one job: looking at a finished thing and judging it against a checklist. That's a kind of work it's much better at. The same model that wrote a mediocre draft a minute ago will reliably tell you why the draft is mediocre when you ask.

Don't take the model's word for it that the draft is fine. Make it argue against the draft. Then make it fix what it just argued.

Critique your own prompts, too

The same pattern works on your prompts before you send them.

"Here's a prompt I'm about to send you. Before you respond to it, critique the prompt itself. What's ambiguous? What's missing? What would make the output better if I added it?"

You get back a list of clarifications. Add them. Then ask the model to run the prompt. Output is consistently sharper.

When NOT to use this

• Brainstorming and ideation. Critique kills divergent thinking. Get your 20 ideas first, then critique. (See Brainstorm and ideate.)
• First drafts you're throwing away. If the whole point is a quick exploratory draft to see if the angle works, skip the critique loop. It's overhead.
• Tasks the model genuinely got right in one pass. Some tasks are easy. Don't add a step for the sake of it.

Common mistakes

• Soft critique prompts. "Could this be better?" — yes, anything could be better. Ask specifically: three biggest problems, listed.
• Vague rewrite prompts. "Make it better" gets you cosmetic edits. "Rewrite addressing the critique above" gets you real changes.
• Trusting the model's "all good." Sometimes the critique pass says the draft is fine. Don't believe it on the first try — push back: "You said it was fine. Find three problems anyway." The model almost always finds them.
• Forgetting to actually run the rewrite. The critique on its own isn't the output. You need step 3.

Where to go next

• Show, don't tell — pair a critique with an example of what good looks like for sharper rewrites.
• Brainstorm and ideate — when not to use critique mode (during divergent thinking).
• Decompose and verify — for long outputs, critique section-by-section instead of all at once.