Make ChatGPT critique its own draft before you ever see the final version

Why this works

This is the draft, critique, revise pattern, sometimes called self-refine prompting, which is documented to measurably improve output quality over asking for a single-pass answer, because a model evaluating a draft it has already externalized as finished text engages a more critical mode than the generation mode it used to produce that same draft. First-pass generation is optimized to produce a plausible, fluent continuation; a critique pass explicitly told to look for what is wrong activates a different weighting than produce something good does, which is why asking for both in the same turn, rather than trusting the first draft, changes the outcome rather than just adding words. Naming specific critique criteria prevents the critique pass from defaulting to generic praise, the most common failure mode of unscoped self-evaluation, where a model politely validates its own work rather than actually stress-testing it. The instruction that finding zero weaknesses is itself suspicious directly targets a documented tendency for models to rate their own output more favorably than an independent judge would, and requiring the pass-2-to-pass-3 mapping at the end makes the revision auditable, so a reader can check that a named flaw was actually fixed rather than the model quietly rewriting for tone instead.