AI Guides › Claude Mastery

The Two-Pass Rule: Draft, Then Ask It To Attack Its Own Draft

By Nigel Guy · 2 min read

The typical workflow stops at the first good-looking draft. Claude produces something fluent and well-organised, it reads as finished, and "reads as finished" quietly gets treated as "is finished." The draft was never asked to survive an actual critique — it just looked like the kind of thing that would.

The rule: the first draft is a starting position, not a result — the second pass, where you make it attack its own work, is where the real quality gets found.

The mechanism: the two passes

  1. Pass one: draft, without self-editing pressure. Ask for the first version plainly, without loading the prompt down with every possible caveat in advance — trying to pre-empt every flaw in the first prompt usually produces something hedged and cautious rather than something genuinely good to critique.
  2. Pass two: switch roles explicitly. Ask it directly to critique the draft it just produced — as a sceptical editor, a rival, a fact-checker, whichever role fits the material — and to name specific weaknesses: what's unsupported, what's generic, what a critical reader would push back on first.
  3. Ask for the critique before the fix. Get the list of problems as its own output, separate from the rewrite. A critique that arrives already bundled with a smoothed-over fix is easy to skim past without actually registering what was wrong.
  4. Decide which criticisms are real before acting on them. Not every self-flagged issue is worth fixing — some are genuine, some are the model being reflexively hedgy. This judgement call stays yours.
  5. Revise once, deliberately, rather than looping the critique step indefinitely — past a certain point, repeated rounds polish phrasing without changing substance, and that's a sign you're done.

What to skip

Skip asking for the critique in the same message as the original draft request — the model needs the actual draft in front of it to critique specifically, rather than critiquing a hypothetical version of what it's about to write. And skip running this twice on genuinely low-stakes, throwaway text; the two-pass rule earns its keep on anything you'll send, publish, or rely on.

What to skip in the critique itself

Skip treating "this is well-written" as a useful critique output — ask specifically for weaknesses, gaps, and unsupported claims, since a general quality judgement tends to default to politeness rather than the pointed version you actually need.

Guardrails

All 751 AI guides · JulieMango plans from £17/mo