Iterating and Refining Conversations

Claude is a conversation, not a vending machine. The first response is a starting point, and the skill the exam measures is knowing which kind of correction to apply next. Some problems are fixed with a short follow-up; some are fixed by editing the original prompt; and some are best fixed by abandoning the thread and starting again with what you learned.

The three ways to iterate

Follow up in the same chat. Use this when the output is broadly right and needs adjustment: "shorter", "more formal", "swap the second and third paragraphs", "now do the same for the German market". Claude keeps the full context of the conversation, so you do not need to repeat the brief.

Edit the original message and resend. In claude.ai you can edit a message you already sent, which regenerates the answer from that point. Use this when the first prompt was missing a block (a constraint, the audience, a fact) and the flaw runs through the whole response. Patching a fundamentally mis-aimed answer with follow-ups often produces a Frankenstein result; fixing the brief produces a clean one.

Start a fresh conversation. Use this when the thread has drifted, when you have changed your mind about the goal, or when the conversation has become so long that Claude is juggling contradictory instructions from earlier turns. Carry forward the best version of the prompt and anything you learned.

Giving feedback that lands

Feedback works best when it is specific and names the gap rather than the symptom. Compare the following.

This isn't right, try again.

Claude has nothing to work with and will produce a variation of the same thing. Instead:

The structure is good. Two changes: the tone is too casual for a board
paper, so remove the rhetorical questions and contractions. And the
risk section omits the regulatory angle we discussed; add one paragraph
on that using the facts in my first message.

Tell Claude what to keep as well as what to change. Without that, it may rewrite the parts you liked.

Asking Claude to critique itself

A reliable iteration move is to ask for a critique before asking for a rewrite: "Before revising, list three weaknesses in the draft above from the point of view of a sceptical CFO." This surfaces issues you had not noticed and lets you choose which to act on. You can then say "fix points one and three, ignore two".

Asking for options is another form of iteration: "Give me three different openings, ranging from conservative to bold." Choosing between options is often faster than describing the one you want.

Recognising conversation drift

Long threads accumulate instructions. Early on you asked for British English; twenty messages later you pasted a US-style template; now the output mixes both. Or you asked Claude to "be brief" for one answer and it has stayed terse ever since. Signs of drift include:

• output that follows an instruction you gave for a different sub-task;
• Claude referring back to a version you have already superseded;
• inconsistent formatting between answers that should match;
• answers that keep getting longer as Claude tries to satisfy everything at once.

When you spot drift, either restate the current rules explicitly ("From now on, ignore the earlier template; use this one") or start a fresh chat with a consolidated prompt. The latter is usually cleaner for anything that will be reused.

Scenario: a job description in four turns

A hiring manager wants a job description. The first prompt names the role, team and three responsibilities. Claude returns a competent draft that is too long and uses a jargon-heavy tone.

Turn two, follow-up: "Cut to 350 words and remove the buzzwords; write the way our engineering blog does, plain and direct."

Turn three, the manager notices salary is missing because they never mentioned it. This is a missing fact, so they edit the original message to add the salary band and the hybrid policy, and regenerate. The new draft includes both in the right places.

Turn four: "Give me two alternative opening paragraphs, one leading with the mission and one leading with the team." The manager picks one and is done. Total time: a few minutes. The key was choosing the right type of iteration at each step rather than typing "try again".

When to stop iterating

Diminishing returns set in quickly. If three rounds of refinement have not got you there, the problem is usually upstream: an unclear goal, missing material or a task Claude is not well suited to. Step back, rewrite the brief using the five blocks, and start clean. The exam rewards candidates who recognise that endless tweaking is a symptom, not a strategy.

Exam focus

• Follow-ups suit small adjustments; editing and resending the original prompt suits missing blocks that affect the whole answer; a fresh chat suits drift or a changed goal.
• Specific feedback names what to keep and what to change; "try again" gives Claude nothing to act on.
• Asking Claude to critique a draft or produce options is a legitimate and efficient iteration technique.
• Conversation drift means earlier instructions are leaking into later tasks; restate the rules or restart.
• Repeated unsuccessful iterations point to a flawed brief, not to a need for more iterations.