What you will be able to do
- State the output format, structure and constraints in the request itself
- Say explicitly how far an instruction applies, and what Claude should do when the source material has no answer
- Use a single well-chosen example to show a format that is hard to describe
- Ask for the depth you need instead of hoping Claude infers it, and know what to do when a response comes back too brief
1.Say what the finished output looks like
A request describes the work. A good request also describes the finished result. Anthropic's prompting guide puts it in one line: be specific about the desired output format and constraints. The prompting blog adds a habit worth building: state what the output should include, not just what to work on. "Look at these supplier quotes" names the work. "Produce a table with one row per supplier, with columns for price, lead time and warranty" names the result.
The blog's meal-planning example shows the difference. The vague version asks for a Mediterranean meal plan. The specific version sets the purpose, a numeric limit, the structure and the level of detail:
Design a Mediterranean diet meal plan for pre-diabetic management. 1,800 calories daily, emphasis on low glycemic foods. List breakfast, lunch, dinner, and one snack with complete nutritional breakdowns.| Element | Examples given in the source | Business-task illustration |
|---|---|---|
| Clear constraints | word count, format, timeline | Must fit on one page |
| Relevant context | who the audience is, what the goal is | For parents, so they know the new pick-up time |
| Desired output structure | table, list, paragraph | Columns matching the tracking spreadsheet the team already uses |
| Requirements or restrictions | dietary needs, budget limits, technical constraints | Only name suppliers on the approved shortlist |
Phrase format instructions as the thing you want. The blog's advice is to tell the model what to do instead of what not to do: rather than "Do not use markdown", write "Your response should be composed of smoothly flowing prose paragraphs." Anthropic's Sonnet 5 guide finds the same thing for length and style: positive examples of the concision you want tend to work better than negative examples or instructions about what to avoid.
A team's prompt says "Do not use markdown in your response," but Claude's output still includes headings and bullet lists inconsistently. What rewording is more effective at controlling the format?
Correct answer: A — State that the response should be composed of smoothly flowing prose paragraphs instead
- A. Correct. Telling Claude what to do instead of what not to do, such as specifying that the response should be smoothly flowing prose, is a more effective way to steer output formatting than a negative instruction.
- B. Repeating the same negative phrasing does not change the underlying issue that negative instructions are less effective than positive ones for controlling format.
- C. Temperature controls randomness in word choice, not adherence to formatting instructions, so it would not reliably fix markdown usage.
- D. Manually stripping markdown after the fact is a workaround rather than a prompting fix, and doesn't address why the instruction isn't being followed.
2.State the scope, and say what happens at the edges
Two kinds of instruction are easy to leave implied, and both cause trouble. The first is scope. Anthropic's Sonnet 5 guide says the model reads prompts literally and explicitly, especially at lower effort levels. It does not silently extend an instruction from one item to another, and it does not infer requests you did not make. The guide calls this literalism a strength for predictable work, and its advice follows from it: if you want an instruction applied broadly, state the scope explicitly. Its own example wording is "Apply this formatting to every section, not just the first one."
The second is what Claude should do when the material you supplied does not contain the answer. If you do not say, the model has to decide for itself. The prompting blog recommends giving explicit permission to express uncertainty rather than guess, and says this reduces hallucinations. Its example request ends with the clause that does this: "If the data is insufficient to draw conclusions, say so rather than speculating." For a request grounded in a pasted handbook, policy or dataset, one sentence about what to do when the source is silent is often the most useful thing you can add.
The request never said what to do when the policy is silent, so Claude has to judge for itself. A plausible-sounding general answer is a real risk. Adding something like "If the policy does not cover the question, say so and suggest contacting HR rather than guessing" gives Claude explicit permission to report the gap instead of filling it.
3.When a format is hard to describe, show one
Some requirements are easier to show than to describe, such as a house style for summaries, the rhythm of a status update, or the exact shape of a row. The prompting guide calls examples one of the most reliable ways to steer Claude's format, tone and structure. The prompting blog lists when to use them: when the format is easier to show than describe, when you need a particular tone or style, when the task involves subtle patterns or conventions, and when simple instructions have not produced consistent results.
Examples carry weight, so choose them carefully. The blog warns that current models pay very close attention to details in examples, so your example should show the behaviour you want and nothing you would rather avoid. The guide says good examples are relevant, meaning close to your actual use case, and varied enough that Claude does not copy an accidental pattern. They should also be wrapped in <example> tags so they are clearly separate from your instructions. On quantity, the blog's tip is to start with one example and add more only if the output still does not match.
4.Ask for the depth you need
The last thing to specify is how thorough the answer should be. The Sonnet 5 guide notes that the model matches response length to the complexity of the task instead of using a fixed length. In practice that means shorter answers to simple lookups and longer ones to open-ended analysis. If a request looks simple, a short answer is a reasonable reading of it, even when you wanted a thorough one.
The fix is to say so. The prompting guide advises that if you want work that goes above and beyond, you should ask for it explicitly rather than rely on the model to infer it from a vague prompt. Its example turns "Create an analytics dashboard" into a request that asks for as many relevant features as possible and a fully featured implementation. The blog's version of the same point is to be specific about quality and depth expectations. The same levers also work in reverse: for a shorter answer, the latency guidance says to ask Claude directly to be concise.
If a response still comes back too short or unfinished, the help centre points to the prompt rather than the model. When responses are too brief or only partly complete, consider reworking your prompts. It also recommends giving Claude the previous prompt and response when you write the next one, so it can continue where it stopped.
Exam traps
Each one states something that sounds right. Open it to see what is actually true.
1.An instruction given for one section or item will naturally be applied to all the others.Why is that wrong?
Claude reads instructions literally and does not silently extend them from one item to another. If a rule should apply everywhere, say so.
Covered in State the scope, and say what happens at the edges
2.The clearest way to control format is a list of things Claude must not do.Why is that wrong?
The documented approach is to describe the output you want. Positive instructions and examples tend to work better than prohibitions.
Covered in Say what the finished output looks like
3.Claude will produce a thorough, comprehensive answer when the task obviously deserves one, even if the request doesn't ask for it.Why is that wrong?
Response length follows how complex the request looks. If you want work that goes beyond the minimum, ask for it explicitly.
Covered in Ask for the depth you need
Sources
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practicesOfficial docs
“Be specific about the desired output format and constraints.”
↩︎ Say what the finished output looks like“Examples are one of the most reliable ways to steer Claude's output format, tone, and structure.”
↩︎ When a format is hard to describe, show one“Relevant: Mirror your actual use case closely.”
↩︎ When a format is hard to describe, show one“explicitly request it rather than relying on the model to infer this from vague prompts.”
↩︎ Exam trap 3 - 2.https://claude.com/blog/best-practices-for-prompt-engineeringSecondary source
“State what you want the output to include, not just what to work on”
↩︎ Say what the finished output looks like“If the data is insufficient to draw conclusions, say so rather than speculating.”
↩︎ State the scope, and say what happens at the edges“Pro tip: Start with one example (one-shot).”
↩︎ When a format is hard to describe, show one“Be specific about quality and depth expectations”
↩︎ Ask for the depth you need“Tell the AI what TO do instead of what NOT to do”
↩︎ Exam trap 2 - 3.https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-sonnet-5Official docs
“Positive examples showing how Claude can communicate with the appropriate level of concision tend to be more effective than negative examples”
↩︎ Say what the finished output looks like“It does not silently generalize an instruction from one item to another”
↩︎ State the scope, and say what happens at the edges“Apply this formatting to every section, not just the first one”
↩︎ State the scope, and say what happens at the edges“Claude Sonnet 5 calibrates response length to the complexity of the task rather than defaulting to a fixed verbosity.”
↩︎ Ask for the depth you need“If you need Claude to apply an instruction broadly, state the scope explicitly”
↩︎ Exam trap 1 - 4.https://support.claude.com/en/articles/8114518-claude-s-response-to-my-prompt-is-too-briefOfficial docs
“If its responses are too brief or only partially complete, consider retooling your prompts.”
↩︎ Ask for the depth you need“We recommend giving Claude the previous prompt and response when writing your next prompt to pick up where it left off.”
↩︎ Ask for the depth you need