What you will be able to do
- Use the Description–Discernment loop to decide what to change after a disappointing response
- Match a symptom (generic, off-target, badly formatted, factually wrong) to the change most likely to fix it
- Tell revisions that add information apart from revisions that only add pressure
- Prompt for the depth you need when a response is too brief or too shallow
Key concept
The Description–Discernment loop — Iterating means you evaluate what Claude returned, work out which part of your request was missing or vague, and then send a sharper description that fixes that gap. Sending the same request again with more force is not iteration.
1.Iteration is a loop, not a retry
Few prompts get exactly what you need on the first try. Anthropic's AI Fluency material describes the working pattern as a loop. You describe what you want, the model responds, you discern what's useful, and you describe again, sharper this time. The key word is *sharper*. Each new round changes the description because of something specific you noticed in the last answer. The same course says the loop is iterative: your first prompt rarely nails it, so you refine it using what each response teaches you.
Description has three parts, and a weak answer usually points to a gap in one of them. Product Description defines the output: what it is, who it is for, and its format and length. Process Description guides the approach: what to look at, which steps to take, and which material to use. Performance Description shapes how the AI behaves while you work together. When you spot a problem, ask which of the three your prompt left vague, and fix that part.
Discernment is the other half of the loop. It means evaluating what came back critically before you decide what to change. You look at the quality of the output, how it was reached, and how the AI behaved. People who skip this step can spend an hour rewording a prompt when the real problem was one that no wording could fix.
Checkpoint 1 of 6· Put it in order
Put the steps of one round of the Description–Discernment loop in order.
- 1.Describe again, sharper, based on what you found
- 2.Describe what you want
- 3.Discern what is useful and what is missing
- 4.Read the model's response
The loop always moves from description to response to evaluation, and the evaluation decides what the next, sharper description contains.
“you describe what you want, the model responds, you discern what's useful, and you describe again, sharper this time.”Source: academy.claude.com
2.Diagnose the failure before you reword
After a weak answer, the first question is not "how do I phrase this better?" It is "what kind of failure is this?" Different failures need different changes. Anthropic's prompt-engineering guidance pairs common symptoms with their fixes:
| What you see | What to change in the next prompt |
|---|---|
| Response is too generic | Add specificity, examples, or an explicit request for comprehensive output; ask it to go beyond the basics |
| Response is off-topic or misses the point | Be more explicit about your actual goal and give context about why you are asking |
| Response format is inconsistent | Add examples of the format you want |
| Task is too complex and results are unreliable | Break it into multiple prompts, each doing one thing well |
| AI makes up information | Explicitly give permission to say "I don't know" when uncertain |
| AI suggests changes when you wanted them made | Be explicit about the action: "Change this function" rather than "Can you suggest changes?" |
Each fix is different. A generic answer needs more specificity. An inconsistent format needs an example. An answer that misses the point needs the goal and the reason behind it. Matching the change to the symptom is what makes the next round better than the last.
Some failures are not wording problems at all. Anthropic's documentation says plainly that not every failing result is best solved by prompt engineering. A common case is output that reads well but contains wrong facts or figures. Better phrasing will not fix that, but grounding will. Give Claude the actual source material, tell it to rely only on that material, and let it say when it does not know. For long documents, ask Claude to extract word-for-word quotes first and base its work on them. Then check the numbers yourself. The AI Fluency guidance is to evaluate outputs whenever accuracy matters.
Analyze this financial data and identify trends. If the data is insufficient to draw conclusions, say so rather than speculating.Checkpoint 2 of 6· Match them up
Match each symptom to the change most likely to fix it.
Tap a term, then the definition that fits it.
Each fix supplies what that particular failure was missing: detail, a model of the format, the purpose, or permission to admit uncertainty.
“Solution: Add specificity, examples, or explicit requests for comprehensive output.”Source: claude.com
Checkpoint 3 of 6· Exam question
After rewriting an instruction to say "write in flowing prose, avoid bullet points," a team still finds Claude's responses returning to bulleted lists in some outputs. They notice their own prompt itself is written using bullet points and headers throughout. Which additional iteration is most likely to help?
Correct answer: A — Remove the bullet points and headers from the prompt itself so the prompt's own style matches the desired flowing-prose output.
- A. Correct. Anthropic notes that the formatting style used in the prompt influences Claude's response style, so removing bullet points and headers from the prompt can help the model adopt the desired flowing-prose style.
- B. Incorrect. Adding more bullet points would reinforce, not reduce, the bullet-point pattern the team is trying to eliminate, making it even harder for Claude to avoid that structure.
- C. Incorrect. The core issue is not instruction recency but the prompt's own style. Moving the anti-bullet instruction to a later turn does not address the underlying influence of the prompt's bulleted formatting.
- D. Incorrect. Increasing temperature affects randomness in word choice and phrasing, but it does not discourage structural elements like lists or bullet points, which are determined more by the prompt's formatting.
3.Changes that help, and changes that only feel like help
Once you know what kind of failure you have, three kinds of change do most of the work.
Add the missing context, including the reason. Claude uses the reason behind an instruction to make related decisions. Anthropic's example compares a bare rule, NEVER use ellipses, with the same rule plus its purpose:
Your response will be read aloud by a text-to-speech engine, so never use ellipses since the text-to-speech engine will not know how to pronounce them.The documentation notes that Claude is smart enough to generalise from the explanation. The reason tells Claude more than capital letters can.
Tighten the specification. Write down the audience, length, structure and constraints you had in mind but left out of the prompt.
Show an example. When the problem is format, tone or structure, an example is one of the most reliable fixes. Start with one example, and add more only if the output still misses. Examples should closely mirror your real use case and vary enough that Claude does not pick up patterns you didn't intend.
Other changes feel productive but do little. Emphasis such as capitals, "IMPORTANT" or repeated warnings tells Claude that a rule matters, but not why it matters or how far it applies. A follow-up like "make it better" gives the model nothing new to work with. The AI Fluency exercise asks instead for one concrete revision prompt that names the change you want. Sending the identical prompt again is a diagnostic, not a repair. Anthropic suggests running the same prompt several times and comparing the outputs to spot inconsistencies, not as a way to fix them. Finally, putting every technique into one revision makes it impossible to tell which change helped.
Checkpoint 4 of 6· Check yourself
A bid writer's reviewers keep disliking how Claude organises long proposal drafts, and each rewrite costs a full round. Which change to the prompt is most likely to fix the organisation?
Structure is easier to show than to describe. An example gives Claude the target organisation, while emphasis, repetition and 'make it better' add no new information.
“Examples are one of the most reliable ways to steer Claude's output format, tone, and structure.”Source: platform.claude.com
4.When the answer is too brief or too shallow
A short or shallow first answer is a common disappointment, and the fix is usually in the prompt. Anthropic's help centre says that if Claude's responses are too brief or only partially complete, you should consider retooling your prompts. The prompting guide also notes that Claude's latest models are generally less verbose and may skip detailed summaries unless you ask for them. If you want depth, you have to ask for it.
The fix is to describe the depth you want instead of hoping for it. Anthropic's own before-and-after example turns a one-line request into a fuller one by adding two sentences: "Include as many relevant features and interactions as possible. Go beyond the basics to create a fully-featured implementation." The same approach works for a briefing or a report. Say who will read it, what they need to decide or do afterwards, which areas need detailed treatment, and roughly how long it should be. If the answer is generic as well as short, add an example of the depth you expect or ask explicitly for comprehensive output.
If a response stops partway through, the help centre recommends including the previous prompt and response when you write the next prompt, so that Claude can pick up where it left off. Starting over usually just gets you another short first attempt.
Checkpoint 5 of 6· Check yourself
A communications officer gets a first draft that is accurate but far too shallow for a leadership briefing. Which follow-up is most likely to improve the second attempt?
Shallow output usually means the prompt never defined the depth it wanted. Naming the audience, the purpose and the areas that need detail tells Claude what 'deep enough' means. The other options add no new information.
“Be specific about quality and depth expectations”Source: claude.com
Checkpoint 6 of 6· Exam question
A prompt asks Claude to solve multi-step math word problems, but a review of outputs shows occasional arithmetic slips that go uncaught. The team wants to iterate on the prompt to catch these errors before the final answer is returned, without switching to a different evaluation pipeline. Which addition to the prompt best addresses this?
Correct answer: A — Append a step instructing Claude to verify its answer against the problem constraints before giving the final answer.
- A. Correct. Adding an explicit self-verification step prompts Claude to check its solution against the original constraints, which can catch arithmetic errors before the final output. This technique directly addresses the need for error detection without altering the evaluation pipeline.
- B. Incorrect. Removing instruction to show steps discourages structured, step-by-step reasoning that helps maintain accuracy on multi-step math problems. Without explicit reasoning, arithmetic slips are more likely to occur and go unnoticed.
- C. Incorrect. Suppressing intermediate work hides the reasoning chain, preventing both the model and reviewers from spotting where arithmetic errors happen. It eliminates any chance for self-correction and does not add a verification mechanism.
- D. Incorrect. Repeating the same approach twice does not introduce a genuine check, because identical reasoning is likely to reproduce the same mistake. Since only the first result is used, the second computation provides no error-catching benefit.
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.If a well-written draft quotes wrong figures, rewording the request more precisely will fix the numbers.Why is that wrong?
Factual errors are a grounding problem. Supply the source material, restrict Claude to it, allow it to say it does not know, and verify the figures yourself.
Covered in Diagnose the failure before you reword
2.Writing a rule in capitals, such as NEVER do X, is the most effective way to make Claude follow it.Why is that wrong?
Explaining the reason behind the rule works better, because Claude can apply the reason to related cases the rule does not mention.
Covered in Changes that help, and changes that only feel like help
3.A short or incomplete answer means Claude has hit a limit, so changing the prompt will not help.Why is that wrong?
Anthropic's help centre treats brief or partial answers as a prompting issue: retool the prompt, and include the previous prompt and response so Claude can continue.
Covered in When the answer is too brief or too shallow
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“The loop is iterative. Your first prompt rarely nails it; refine using what each response teaches you.”
↩︎ Iteration is a loop, not a retry“Evaluate outputs when accuracy matters”
↩︎ Diagnose the failure before you reword“Write one concrete revision prompt”
↩︎ Changes that help, and changes that only feel like help“you describe what you want, the model responds, you discern what's useful, and you describe again, sharper this time.”
↩︎ Key concept“you describe what you want, the model responds, you discern what's useful, and you describe again, sharper this time.”
↩︎ Checkpoint - 2.
“It includes Product Description (defining outputs), Process Description (guiding the approach), and Performance Description (shaping AI behavior during collaboration)”
↩︎ Iteration is a loop, not a retry - 3.
“Not every success criteria or failing eval is best solved by prompt engineering.”
↩︎ Diagnose the failure before you reword - 4.https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinationsOfficial docs
“ask Claude to extract word-for-word quotes first before performing its task.”
↩︎ Diagnose the failure before you reword“Run Claude through the same prompt multiple times and compare the outputs.”
↩︎ Changes that help, and changes that only feel like help“Explicitly instruct Claude to only use information from provided documents and not its general knowledge.”
↩︎ Exam trap 1“Explicitly instruct Claude to only use information from provided documents and not its general knowledge.”
↩︎ Prediction - 5.https://claude.com/blog/best-practices-for-prompt-engineeringSecondary source
“Be more explicit about your actual goal.”
↩︎ Diagnose the failure before you reword“Start with one example (one-shot).”
↩︎ Changes that help, and changes that only feel like help“Solution: Add specificity, examples, or explicit requests for comprehensive output.”
↩︎ Checkpoint“Be specific about quality and depth expectations”
↩︎ Checkpoint - 6.https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practicesOfficial docs
“Providing context or motivation behind your instructions, such as explaining to Claude why such behavior is important, can help Claude better understand your goals”
↩︎ Changes that help, and changes that only feel like help“Relevant: Mirror your actual use case closely.”
↩︎ Changes that help, and changes that only feel like help“Less verbose: May skip detailed summaries for efficiency unless prompted otherwise”
↩︎ When the answer is too brief or too shallow“Include as many relevant features and interactions as possible. Go beyond the basics to create a fully-featured implementation.”
↩︎ When the answer is too brief or too shallow“Claude is smart enough to generalize from the explanation.”
↩︎ Exam trap 2“Examples are one of the most reliable ways to steer Claude's output format, tone, and structure.”
↩︎ Checkpoint - 7.https://support.claude.com/en/articles/8114518-claude-s-response-to-my-prompt-is-too-briefOfficial docs
“We recommend giving Claude the previous prompt and response when writing your next prompt to pick up where it left off.”
↩︎ When the answer is too brief or too shallow“If its responses are too brief or only partially complete, consider retooling your prompts.”
↩︎ Exam trap 3