1.3 Iterating Deliberately on Claude's Output

Prompting is iterative by nature -- the first response rarely arrives exactly right, and that's expected, not a failure. The skill this task statement tests is what you do next. Iteration means diagnosing exactly why the output missed and making one deliberate, targeted change. It does not mean hitting "regenerate" and hoping the next roll of the dice lands better.

This distinction matters because regeneration and iteration can look identical from the outside -- both produce a new response -- but they work completely differently. Regeneration resamples the same underspecified prompt and hopes for a better draw; it fixes nothing about the actual gap. Iteration identifies the gap first, then closes it. Only one of these is a repeatable skill; the other is luck.

Deliberate iteration follows three steps, always in this order. First, **read the output against your intent** -- identify exactly what's wrong: wrong format, missing detail, wrong tone, a factual error, too long, too short. Second, **make a targeted change** -- add the missing constraint, supply an example, tighten the audience, or split the task into steps. Third, **re-run and compare** -- keep the better version, and repeat the loop if a gap remains.

• 1.Read -- diagnose specifically what's wrong: format, detail, tone, accuracy, or length.
• 2.Change -- make one targeted edit that addresses that specific gap.
• 3.Recheck -- re-run, compare against the prior version, and keep the better one.

The discipline that makes this loop actually work is isolating one variable per iteration. If you rewrite the tone, the format, the length, and the constraints all in the same edit, and the new output is better, you have no way of knowing which change actually helped -- and if it's worse, you don't know which change to undo. Change one thing, observe the effect, and only then decide whether to layer on a second change.

This is slower in the moment but faster overall, because it builds a reliable mental model of what each lever does. An associate who has iterated this way a dozen times knows, with confidence, that adding a length constraint fixes overly long drafts and that adding an example fixes tone mismatches -- because they tested each one individually rather than changing five things at once and guessing at the cause.

This loop is the same pattern as the *Diligence* competency in Anthropic's AI Fluency framework: you stay responsible for judging and refining the result, rather than treating the first output as automatically final, and rather than outsourcing that judgment back to the model by asking it to grade its own work.

Framed this way, iteration isn't a workaround for Claude's limitations -- it's the ordinary, expected shape of getting useful work out of any capable collaborator, human or AI. A first draft is a draft. The associate's job is to read it critically, name the gap, and close it -- and that responsibility doesn't transfer to the model just because the model produced the draft.

Task Statement 1.3 questions typically present a scenario where a first output missed the mark and ask for the best next step, or ask you to identify why a described approach isn't real iteration.

• 1.Repeatedly regenerating the same prompt without changing anything -- this doesn't diagnose or fix the actual gap.
• 2.Changing five things in one edit (tone, format, length, and constraints together), so you can't tell what helped or hurt.
• 3.Treating a first draft as final without reading it critically against your original intent.
• 4.Asking Claude to judge or grade its own output instead of doing that diagnosis yourself.

The correct answer to "the output missed the mark, what's next?" is essentially always some version of: diagnose the specific gap, make one targeted change, and recheck -- never "try again" without a diagnosis, and never "change everything at once."