Task statement 1.3 · Prompting and task execution
1.3 · Iterating Prompts
In Wrenfield Housing, the course example, the communications team writes the spring tenant newsletter with Claude. The first draft is close but not right, and each reply the team sends either moves the draft towards what tenants need or only makes it different.
share
The tempting reply to a disappointing draft is "make it better", because it is quick and leaves the details to Claude. Claude cannot tell which part fell short, so it changes several things at once, and the parts that were already right often change with the rest. The opposite habit fails too. Rewriting the whole prompt after every draft throws away what was working, and when several things change at once, nobody can tell which change helped.
KnowledgeThe iteration loop
To iterate a prompt is to improve output quality over several rounds: you read what Claude produced, compare it with what you need, and send a change. Anthropic's prompt engineering overview names three things to have before you start improving a prompt: a clear definition of success, a way to test the result against it, and a first draft. For a business document, the definition of success can be a short list of checks, meaning tests that a good result passes or fails. Anthropic's guide to success criteria asks for specific criteria, so for a newsletter "400 words or fewer" is a check, while "punchy" is not. With checks written down, every round has a target, and any two drafts can be compared on the same test. This rules out any answer that keeps asking for "better" without saying better at what.
The loop then runs in five steps, and the first one happens before you reply to the first draft:
- Define better: write down two to four checks the result must pass.
- Read against the checks: note which checks the draft fails, not only how it feels.
- Send one change: name the change, the reason for it, and what to keep.
- Compare: judge the new draft against the same checks as the last one.
- Stop or restart: stop when every check passes, or start again when the problem began in the prompt.
Interactive · one change per turn, judged against three checks
Step through the newsletter's four turns and see which checks each draft passes.
Turn 1: the prompt gives the three items, the readers and “one page”. Draft 1 runs to 650 words, puts the boiler item last and never tells tenants what to do.
Turn 2 asks for one change: end each item with a “What to do:” line. Draft 2 passes that check and grows to 700 words, and the boiler item is still last.
Turn 3 moves the boiler item first, because tenants must act before 3 March, and keeps everything else. Draft 3 passes two checks, and only the length is left.
Turn 4 cuts the draft to 400 words or fewer and keeps the dates and actions. Draft 4 passes all three checks, so the team stops.
A follow-up turn works because Claude reads the whole conversation, so a reply can point at "the boiler item" or "everything else" without repeating the prompt. The Claude Help Center lists giving feedback as one way to improve an answer: follow-up instructions, clarifications, or a request to rewrite. Feedback is specific when it names the one change, the reason for it, and what to keep. The reason lets Claude apply the change sensibly, and the "keep" line protects the parts that already pass. Anthropic's prompting advice is to test each addition to see whether it improves the result. You can only do that when you know which message caused which difference, which is why you send one change per turn.
KnowledgeFollow up, or revise the prompt and start again
You can send a change in two ways: reply in the same chat, or rewrite the original prompt with what you learned and send it in a new chat. A reply keeps everything that came before, which helps when the draft is mostly right. It also keeps the drafts that went wrong, because Claude reads every earlier turn along with your new message. Sometimes the problem started in the original prompt, such as a missing reader, a wrong fact or a missing item. A revised prompt then fixes it at the source, and the new chat no longer carries the drafts that went wrong. A revised prompt also suits a task that repeats, because the checks you learned go straight into next time's prompt:
Follow-up turn
Use it when the draft is mostly right and only one part falls short.
Carries forward: every earlier turn, including the drafts that went wrong.
Revised prompt
Use it when the problem began in the prompt, or when the task repeats.
Carries forward: only what you write into the new prompt, so add what you learned.
SkillWorked example: the spring newsletter
Wrenfield's spring newsletter goes to every tenant household and has three items: the Millbrook boiler servicing from 3 to 14 March, the repairs line's new opening hours, and the spring tenants' forum. Draft 1 ran to 650 words, put the boiler item last and never told tenants what to do. Before replying, the team wrote down three checks: each item ends with what to do, the item that needs action soonest comes first, and the newsletter fits its one-page layout at 400 words or fewer. The first reply was vague:
# turn 2, before Make it better and more engaging. → 720 words, three exclamation marks, boiler item still last # turn 2, after Good start. One change: end each item with a line that starts "What to do:", so tenants know their next step. Keep everything else as it is. → every item ends with an action; nothing else moved
The bad round came from feedback that named no check, so Claude guessed that "better" meant livelier and changed the tone instead of the structure. The specific reply named one change and its reason, and its last line held the rest of the draft still. Two more turns followed the same pattern, as the figure shows, and draft 4 passed all three checks, so the team stopped there. The communications lead still read it before it went out, as Wrenfield's AI policy requires. The team then wrote the three checks into the prompt for the summer issue.
Stop when the checks pass. A turn sent after every check passes, such as "make it a bit sharper", has no check to judge it by, so it can undo a check that already passed. Keep the draft that met the checks, and leave final polish to the person who reviews it.
- Write down what "better" means as checks before you reply, so every round has a target and drafts can be compared.
- One change per turn, with its reason and what to keep, beats "make it better", because Claude cannot guess which part fell short.
- Follow up when the draft is mostly right; revise the prompt in a new chat when the problem started in the prompt.