Iteration over the first draft

A workflow that evaluates an initial AI output, identifies a specific failure, and runs another pass with better evidence, criteria, or instructions. It stops when the output meets the task's checks or another pass is no longer worth its cost.

Iteration over the first draft treats the initial output as evidence about the task. A project update may omit a metric, hide a blocker in vague language, or use the wrong level of detail. Compare it with the source and the acceptance criteria, name the failure, and revise that part with the missing evidence. The next pass is useful because its input changed, not because later drafts are automatically better. A clean first result can pass immediately; a flawed fifth result should still fail.

Builder example

A production loop needs an observable reason to continue. A summarization tool can check required fields and source citations, route only failed sections through revision, then rerun the same checks. A draft-email feature can compare tone and factual claims with examples and account data. Low-risk outputs may publish after automated checks, while higher-risk or ambiguous failures can route to a person or another specialist. The stopping rule belongs in the workflow rather than in a universal number of passes.

Common confusion: Repeated generation is not improvement by itself. The next pass needs new information: a failed check, a missing fact, a clearer criterion, a useful example, or a different method. Self-critique can suggest problems, but an independent source or executable check gives stronger evidence when one is available.