Home › Study guides › CCAO-F › Domain 1 › Lesson 1.3

CCAO-F · Domain 1 · 14% of the exam · Lesson 1.3 · 19 min read

Iterating on a prompt, one change at a time

Why Claude's first answer is only a first draft, and how to improve a prompt on purpose: set criteria, name the defect, change one thing and retest.

Written against objective 1.3 of the official CCAO-F exam guide (Version 1.0, effective July 2026). An independent resource, not affiliated with Anthropic; the practice questions are written from scratch.

1.3.1 Why the first answer is only a first draft

It's early March. You're the marketing specialist at a regional chain of twelve fitness studios. You need a spring email to win back lapsed members: people whose membership ended in the past year. You type "Write a spring email to win back members who cancelled", and Claude returns a friendly, polished draft in seconds.

Then you read it properly. It opens "Dear Valued Member", as if the reader had never left. It offers 20% off, a free personal training session and a bring-a-friend pass, none of which you approved; your one real offer is a half-price first month. It runs to about 240 words and closes with "Don't let another season slip away", a guilt trip your brand never uses. It would suit a gym somewhere. It doesn't suit yours.

What most people do next is reply "make it better" or "more exciting". Claude obliges with a new draft: different words, some of the same problems, a few new ones. Four rounds later you have four drafts and no idea whether any of them is closer. Nothing is wrong with Claude here. Each round gave it no new information, and gave you no standard to judge by.

The fix is to iterate deliberately. Anthropic's Claude 101 course puts the mindset in one line: treat first drafts as starting points. Iteration means improving a prompt, your brief to Claude, on purpose. Decide what good looks like before you judge, compare the draft with it and name the specific defect. Then change the prompt to fix exactly that, and run it again on the same input: the same case as before, here the same group of lapsed members.

1.3.2 Decide what good looks like before you judge

Here is the question that trips people up: how do you know draft 2 is better than draft 1? Without a standard, you judge by feel, and feel rewards whatever is newest or most polished. The fix is a short list of success criteria, written before you run the prompt, saying what the output must do. They belong in the prompt itself, so Claude can aim at them. In iteration they do a second job: they are the test every draft has to pass.

For the win-back email, five criteria are enough. What makes them useful is that each one is a yes-or-no check you can do in seconds, not an impression.

Look at the right-hand column. Once criteria exist, the vague feeling that the draft is "generic" turns into five named defects, and each defect points to its own fix. Two of them, the audience and the offer, come from facts Claude was never given, so no amount of rewording can repair them.

1.3.3 Name the defect, then ask for exactly that

Once you know what is wrong, you have to tell Claude. The tempting reply is "make it better" or "make it more professional". It feels like an instruction, but it carries almost no information. Claude doesn't know which of the many things you disliked, so it changes several of them, and possibly not the one you meant.

Think of a hairdresser. "Make it look nicer" leaves them to guess, and you may walk out with a different cut you like no more than the first. "Take two centimetres off the back and leave the fringe alone" can be carried out and checked. Specific feedback works the same way. It says WHERE the problem is, what is wrong with it, and what to do instead. Claude 101 draws the same contrast: "make it shorter" is fine, but naming what to cut and how to end is better.

For the win-back draft, "make it warmer" becomes: "Delete the last line about letting another season slip away; we never make people feel guilty about missed workouts. End instead with one friendly sentence inviting them to rejoin by 30 April." It names the line, the problem and the replacement. It also gives the reason, so Claude can avoid other guilt-tinged phrases you didn't list.

1.3.4 Follow up in the chat, or fix the brief

Now, where do you make the change? You have two places: a follow-up message in the same chat, or the original prompt itself. They feel interchangeable, but they aren't. A follow-up changes this draft. A fix to the brief changes every draft the prompt will ever produce, including next spring's.

When the draft is basically right and needs a small adjustment, send a follow-up, such as "Move the offer into the first sentence and shorten the second paragraph". When the defect comes from something the brief never said, such as who the readers are, what the offer is or how your brand sounds, fix the brief and run it again.

The way to rerun a fixed brief that works everywhere is to paste the revised version into a new chat. Some versions of the Claude app also let you edit your earlier message instead. Claude then answers the edited message in a separate version of the chat, and the earlier version is kept. Either way, the new draft starts from the corrected brief, not from a patched draft.

Sometimes the thread itself is the problem. After many rounds of corrections, a conversation fills with old drafts, reversed instructions and abandoned ideas, and all of it is still in front of Claude. When old and new start to mix, follow Claude 101's advice: sometimes it's faster to open a new chat with a clearer prompt than to try to redirect.

1.3.5 Let Claude critique the draft and the prompt

You don't have to find every defect alone. Claude can check a draft against your criteria, and it can suggest how to improve the prompt that produced it. Anthropic's AI Fluency course calls the second move a "secret weapon": ask the AI itself to help improve your prompt.

A critique is useful when it is tied to your criteria. In the request below, look at the instruction to quote the failing line, which makes every claim quick to check. Then look at the last instruction, which turns the critique into changes to the prompt. The labels in angle brackets only mark where each pasted text starts and ends.

Below are my success criteria for a spring win-back email, then the latest draft, then the prompt I used.
Criteria: 1. Speaks to people whose membership has ended, not current members. 2. One offer only: half-price first month, no joining fee, rejoin by 30 April. 3. Under 120 words. 4. Warm and plain; never makes readers feel guilty about missed workouts. 5. One call to action: rejoin by 30 April.
Check the draft against each criterion in turn. For each one, say pass or fail, and for a fail, quote the line that fails. Do not rewrite the email.
Then read my prompt and tell me what it is missing that led to each failure, and suggest the wording I should add.
<draft>(paste the latest draft)</draft>
<prompt>(paste the prompt you used)</prompt>

Notice what this is not. "Rate this draft out of 10" is a different request. A score gives you nothing to change and nothing to check, and a confident rating is not evidence. Treat it like notes from a second reviewer: useful, but you decide what to keep, change or discard, because the final output is yours.

1.3.6 Change one thing, retest, keep what works

It's tempting to add the audience, the offer, a voice example and a new length limit in one round. Then if the draft improves, you can't tell which change did the work; if something breaks, you can't tell which change caused it. A cook adjusting a soup adds salt and tastes, then adds lemon and tastes again. A prompt works the same way. Make one or two changes per round, keep the previous version, and compare the two drafts side by side against the criteria. Anthropic's prompting advice says the same: test each addition to see whether it actually improves the result.

Claude's answers also vary from run to run: send the same prompt twice and you can get two different drafts. So a new draft will look different whether or not your change helped. Different is not better: what matters is whether it meets more criteria than the last one. That's also why you rerun on the same input; change the prompt and the input together, and you can't separate the two effects.

Like the others, V4 answers a named defect. V3's body met all five criteria, but its subject line, "Your membership has expired", read like a bill and failed the brand-voice check. So V4 asks for three subject lines in the brand voice, and you pick one.

A prompt that passes on one input may only fit that input, so the last check uses a representative second one: a different case of the kind the prompt will really meet. Here that is a second segment, or group of members: new-year joiners, who signed up in January and cancelled within weeks. For them, V4 mentions "your favourite classes", a detail borrowed from the voice example, which was written for long-standing members. One added line fixes it: match the example's voice but not its details. Rerun both segments; both pass.

Then keep what works. Save the brief in a shared document, with a note on what each part fixed, so you or a colleague can reuse it next season. For work you repeat, it can later become the instructions of a Project, a workspace in Claude that keeps instructions and files together for recurring work. In the saved brief below, look at the segment line, which makes it reusable, and the last line, which came from the retest.

Write a spring win-back email for one segment of our lapsed members: people whose membership has ended. Write to them as former members, not current ones.
Segment: [who they are, for example: joined in January, cancelled within weeks]
Context: we run twelve fitness studios. Spring is when people who left are most likely to return, and we want them back before the summer timetable starts.
Offer (the only one; use it exactly): first month back at half price, no joining fee, for anyone who rejoins by 30 April.
Voice: warm, upbeat and plain, like the example below. Never make readers feel guilty about missed workouts; people come back when they feel welcome, not judged.
Output: three subject lines to choose from, then the email body in under 120 words, ending with one call to action: rejoin by 30 April.
<example>(last autumn's welcome-back email, pasted in full)</example>
Match the voice of the example but not its details: take every fact from this brief, not from the example.

1.3.7 The exam traps

Every trap here is a way of changing something without knowing why it failed or whether the change helped.

• ✗ Replying "make it better" or "more professional". ✓ Say where the problem is, what is wrong and what to do instead. Vague feedback produces a different draft, not a better one.
• ✗ Accepting a new draft because it reads better. ✓ Judge it against the criteria you wrote before the first run, on the same input. Two runs can differ even when nothing has improved.
• ✗ Changing five things in one round. ✓ Change one or two, keep the previous version and compare, so you know what helped and what broke.
• ✗ Sending the same follow-up corrections every time you reuse a prompt. ✓ When the brief lacked something, fix the brief and rerun it. Follow-ups are for small adjustments.
• ✗ Switching to a more capable model because the draft missed information you never gave. ✓ Iterate the prompt first. No model knows your offer or your audience until you provide them.
• ✗ Asking Claude to score its own draft and trusting a high score. ✓ Ask for a critique against your named criteria, with the failing lines quoted, then check it yourself.

1.3.8 Put it together: improve a prompt in deliberate rounds

You now have the whole loop. Criteria come first, then a named defect, a fix in the right place and one change at a time, with a retest on the same input and then on a representative one. The quickest way to believe it is to watch a good prompt lose one ingredient.

The rest of Domain 1 builds on this loop. Prompting by task type (1.4) shows how the criteria and the brief shift between analysis, research, drafting and brainstorming. In Domain 5, a prompt that has proved itself moves into a Project's instructions, so the whole team starts from it. And when a workflow keeps underperforming despite careful iteration, Domain 7 goes further: diagnosing the root cause (7.1) and adjusting your approach from feedback and results over time (7.2).

Key takeaways

• ✓ The first output is a first draft; improve it on purpose rather than by asking again and hoping.
• ✓ Write specific, checkable success criteria before the first run, and judge every draft against them, not by feel.
• ✓ Specific feedback names where the problem is, what is wrong and what to do instead; "make it better" gives Claude nothing to act on.
• ✓ Use a follow-up for small adjustments; when the brief lacked information, fix the brief and rerun, and start a new chat when a thread is muddled.
• ✓ Ask Claude to critique the draft against your criteria or to suggest what the prompt is missing, then check its notes yourself.
• ✓ Change one or two things per round, keep the previous version, and retest on the same input, because different is not better.
• ✓ Retest on a representative second input before you save the prompt for reuse.

Check your understanding

4 questions written for this lesson, then one from the CCAO-F question bank on the same topic. Every answer option is explained, including the ones you did not pick. Nothing is stored.

Sources

• Getting better results (Claude 101)
• The Description-Discernment loop (AI Fluency)
• Effective prompting techniques (AI Fluency)
• Prompt engineering best practices
• What are artifacts and how do I use them?
• Claude Cowork and chat are one Claude
• The CCAO-F exam guide, summarised