What you will be able to do
- Identify the context Claude does not have by default and supply it in the prompt
- Plan a prompt from purpose toward wording: objective, audience, source material, then assembly
- Lay out role, context, source documents, task and output format so Claude parses them unambiguously
- Choose an output format and constraints that match how a specific audience will use the result
- Add one or more examples when a format or tone is easier to show than to describe
Key concept
Context Claude lacks by default — Claude starts every prompt without your goal, your audience, your organisation's norms or any facts that came after its training. Anything the first response depends on that only you know has to be written into the prompt, or the response will not use it.
1.What Claude does not know when you start
Prompt construction means turning a business need into instructions that produce a usable first response. Most of the work is working out what Claude cannot know unless you tell it. One useful picture is a capable new hire on their first morning. They are skilled, but they don't know your customers, your house style, why the report matters or who will read it. Anthropic uses the same picture, a newly hired contractor with no context about you, your task or your organisation.
Three kinds of knowledge are missing by default. The first is the purpose: what decision or action the output feeds. The second is the audience: who reads it, what they already know and how they will read it. The third is your source facts: product specs, policies, contracts, interview notes and anything recent. What Claude knows comes from training data and stops at a knowledge cutoff. Your internal documents and last month's policy change are not in it, so they have to be in the prompt.
A practical test catches most gaps before you send anything. Picture a colleague who knows nothing about the task reading your prompt and trying to follow it. If they would need to ask "for whom?", "in what format?" or "based on which document?", Claude has the same gap. Claude will still answer, though, and it fills the gap with a plausible default instead of asking.
The purpose and the audience. Claude doesn't know the reader is the most important client, that they are angry, or what the email needs to achieve, for example keeping the account and giving a credible new delivery date. Without those, "better" can only mean generic polish.
Checkpoint 1 of 9· Check yourself
An operations manager types into Claude: "Draft a reminder to staff about our updated expense policy, which changed last month." They do not paste the policy. What is the most likely result?
Claude's knowledge comes from training data and stops at a cutoff, so an internal policy that changed last month is not in it. Claude fills the gap with a plausible default instead of asking. The policy text has to be in the prompt.
“What generative AI knows comes entirely from training data and is frozen at the knowledge cutoff.”Source: academy.claude.com
2.Work from purpose toward wording
Strong prompts are planned in reverse. You decide what the reader must achieve before you gather any material or write any instructions. One way to think about it is a chain that runs from the user's need, to a requirement, to the instruction Claude acts on. Prompt wording is only the last link. If an earlier link is wrong, polishing the wording cannot fix it, because the AI only works from what you tell it.
Take a team lead building a customer-facing FAQ from an internal product spec. The planning order is:
- Define the outcome: what customers should be able to do after reading, such as complete a return without contacting support.
- Describe the readers: what they already know and the words they use, so the answers start where they are.
- Gather the facts Claude cannot know on its own, here the product spec and the return-policy pages.
- Write the prompt, combining objective, reader description, source documents, task and a required answer format.
Gathering documents first looks efficient, but you can't tell which material matters until the goal is fixed. A spec has dozens of sections. Only the goal "complete a return without support" tells you the return policy is essential and the engineering tolerances are noise. The same goes for wording. Every adjective in the plan should be a decision you can defend. "Clear answers" becomes "answers a first-time buyer can follow on a phone without other help".
A goal is different from a format. "Convince my team this timeline is realistic" is a goal, and "three bullet points" is a format. A prompt that gives only the format can be followed to the letter and still be useless. "Make this shorter" can produce a shorter draft that loses the one finding the executive needed. Put the goal next to the instruction and Claude can work toward what you meant, not only what you said.
Checkpoint 2 of 9· Check yourself
A finance analyst asked Claude to "make this report shorter" before it goes to the CFO. The draft that came back is shorter, but it has dropped the one cost finding the CFO needed. Which revision to the prompt fixes the cause?
"Make this shorter" is a format instruction. Claude can follow it to the letter and still produce something useless. Adding the goal (what the reader must get from the report) lets Claude work toward what was meant. A word limit, louder wording or more material is still not a goal.
“Then re-prompt with the goal stated explicitly alongside the instruction”Source: academy.claude.com
Checkpoint 3 of 9· Exam question
A team lead will build a prompt for a customer-facing FAQ from an internal product spec. Which sequence of steps should the lead follow, from the first preparation step to the finished prompt?
Correct answer: B — Define what customers should achieve → List customer knowledge and vocabulary → Gather product spec and policy pages → Write the prompt with task and format
- A. Listing what customers know before defining what they should achieve puts the readers ahead of the objective. The objective is meant to drive every later choice, so the reader description would be written without a goal to shape it.
- B. The objective is set first because it drives every later choice. Next come the readers' knowledge and vocabulary, then the sources Claude cannot know on its own. The prompt is written last, once all of those ingredients exist.
- C. Gathering the spec and return policy pages before describing the readers means the sources are picked without knowing the readers' starting point and terms. The reader description should shape which material is selected and how the answers are phrased.
- D. Starting with the readers and gathering sources before the objective leaves both steps with no purpose to guide them. The objective has to come first so that the reader description and source selection both serve it.
3.The parts of a prompt and where they go
Once the plan is settled, the prompt itself has a small set of parts. Each part gives Claude something it can't infer. Not every prompt needs all of them, but a first prompt for business work usually needs most.
| Part | What it supplies | Example |
|---|---|---|
| Role | Focuses behaviour and tone for the use case | "You are a helpful coding assistant specializing in Python." |
| Context | Goal, audience, and why constraints exist | Leadership will read this before a funding decision |
| Source material | Facts Claude does not have, kept apart from instructions | Report wrapped in <document> tags with a <source> subtag |
| Task | The action, led by a direct verb | Identify the three findings that most affect the decision |
| Output format and constraints | Shape, length and structure of the answer | One-page memo, under 300 words, recommendation first |
A role can be a single sentence, and it still makes a difference. In API use it goes in the system prompt, separate from the user's request:
client = anthropic.Anthropic()
message = client.messages.create(
model="claude-opus-5-5",
max_tokens=1024,
system="You are a helpful coding assistant specializing in Python.",
messages=[
{"role": "user", "content": "How do I sort a list of dictionaries by key?"}
],
)Order matters when the source material is long. Long documents belong near the top, above the query, the instructions and any examples. Putting them there improves performance for large, data-rich inputs. Picture a project manager summarising a long research report for leadership. A layout that follows this rule is:
- The business goal, including that leadership reads the result before a funding decision.
- The full report inside descriptive tags.
- The request: identify the three findings that most affect the funding decision and explain why.
- The format: a one-page memo under 300 words with the recommendation in the opening line.
XML-style tags stop Claude from confusing your instructions with the material they refer to. This matters most when a prompt mixes instructions, context, examples and variable inputs. Give each kind of content its own consistently named tag. For several documents, nest each one inside a parent tag with its source as metadata, then put the task after them:
<documents> <document index="1"> <source>annual_report_2023.pdf</source> <document_content> {{ANNUAL_REPORT}} </document_content> </document> <document index="2"> <source>competitor_analysis_q2.xlsx</source> <document_content> {{COMPETITOR_ANALYSIS}} </document_content> </document> </documents> Analyze the annual report and competitor analysis. Identify strategic advantages and recommend Q3 focus areas.When the work needs several dependent passes, such as extracting first and then drafting, a single prompt is often the wrong tool. That is covered under breaking work into ordered steps.
Checkpoint 4 of 9· Fill the gap
This API call gives Claude a one-sentence role, kept separate from the user's request. Which parameter name completes it?
client = anthropic.Anthropic()
message = client.messages.create(
model="claude-opus-5-5",
max_tokens=1024,
? ="You are a helpful coding assistant specializing in Python.",
messages=[
{"role": "user", "content": "How do I sort a list of dictionaries by key?"}
],
)In API use the role goes in the system prompt, passed as the system parameter, separate from the user's request in messages. Even a single sentence there focuses Claude's behaviour and tone for the use case.
Checkpoint 5 of 9· Exam question
A project manager is assembling a first prompt asking Claude to summarise a long research report for a leadership team that will read it before a funding decision. Which sequence puts the parts of the prompt in the most effective order?
Correct answer: A — Explain business goal and readers → Paste report inside descriptive tags → Ask for three key findings → Set memo format and length
- A. The goal and readers frame everything, the tagged report follows so it is clearly separated from the instructions, the specific question comes after Claude has the material, and the format and length requirements close the prompt where they are freshest when writing begins.
- B. The report is pasted before Claude knows the business goal or that leadership will read the result before a funding decision. Claude has to read a long document without the context that tells it what matters, so the framing arrives too late to shape how it reads the material.
- C. The request for the three key findings comes before the report is included. Claude is asked a question about material it has not yet seen, instead of reading the report first and then answering a question placed after it.
- D. The request for findings is placed last, after the format requirements. The memo format and length are stated before Claude knows what it is being asked to produce, and the final instruction is no longer the task itself, so the format loses its place at the end where it is freshest as writing begins.
Sources3
4.Choosing an output format for the reader
The output format is where the audience analysis pays off. A format is right when it matches how the reader will use the output, not when it looks thorough. State the shape (table, list or paragraphs), the length and the order. If you leave them out, Claude picks a sensible default for an unknown reader, and that default is rarely right for your actual reader.
Three choices turn an audience description into a format:
- Filter to what the reader asked about. In the contract case, that means only clauses that change cost or deadlines.
- Order by the reader's priority, here the size of the money or timing impact, not the order of the source document.
- Cap the length to fit how they read, here a few short items they can take in between meetings.
The tempting alternatives each fail one of these tests. A clause-by-clause table in contract order covers everything, so the reader has to find the few rows that matter. A two-paragraph narrative holds the cost and deadline points back until the second paragraph, which is the part a busy reader is least likely to reach. A memo with headings for parties, definitions and obligations follows the contract's structure, not the reader's question. Each of these is a good format for some reader, just not this one.
Constraints work better when they say what to do and why. "NEVER use ellipses" is a bare rule. "Your response will be read aloud by a text-to-speech engine, so never use ellipses" gives the reason. Claude can generalise from that reason to related cases, for example other symbols a speech engine would stumble on. In the same way, a prompt that keeps getting bullet points works better if it asks for "smoothly flowing prose paragraphs" than if it repeats "don't use bullet points". A positive instruction gives Claude a target. A prohibition only rules one thing out.
Checkpoint 6 of 9· Check yourself
Frontline support agents will glance at Claude's summary of a 30-page release note in the middle of a call, when a customer asks whether their bug is fixed. Which format instruction fits these readers best?
The right format depends on how the output will be used. Only this option keeps just what agents need on a call, puts it in the order of their priority and keeps it short. The other options follow the source's structure, pad the answer with changes customers never see, or hold the fixes back until the second paragraph.
“Describing how the output will be used”Source: claude.com
Checkpoint 7 of 9· Check yourself
Claude keeps answering in bullet lists, but the text is pasted into a formal letter to a client. Which instruction is most likely to fix this?
A positive instruction gives Claude a target, while a prohibition only rules one thing out. Giving the reason (a formal letter) also lets Claude generalise to other formatting choices that would not suit the letter. Repeating or moving the prohibition still leaves Claude without a target.
“Instead of: "Do not use markdown in your response" Try: "Your response should be composed of smoothly flowing prose paragraphs"”Source: claude.com
Checkpoint 8 of 9· Exam question
An operations manager must brief executives on a 12-page vendor contract. The executives read on their phones between meetings and only want to know what changes costs or deadlines. Which output-format instruction best fits this audience?
Correct answer: B — Ask for five bullets of under 20 words each, ordered by money or timing impact, plus one final sentence naming the largest risk.
- A. Incorrect. One row per clause in contract order forces readers to scan clauses that change nothing about money or timing, and a multi-column table is awkward on a phone screen.
- B. Correct. Five short bullets ordered by money and timing impact, plus one sentence on the largest risk, give phone readers between meetings only what they need.
- C. Incorrect. Dense narrative paragraphs are slow to read on a phone, and putting the pricing and deadline points in the second paragraph buries the only information these readers want.
- D. Incorrect. Headings for parties, definitions and obligations follow the contract's own structure rather than its cost and deadline effects, and a paragraph under each heading makes a long document to read on a phone.
5.Showing the format with examples
Some requirements are easier to show than describe, such as a particular summary voice, a house template or a tone. A few well-chosen examples, called few-shot or multishot prompting, are one of the most reliable ways to set format, tone and structure. They also improve accuracy and consistency. One example of the summary style you want, followed by "now summarise this article in the same style", often does more than a paragraph describing that style.
Claude pays close attention to examples, including details you didn't mean to teach. That makes it worth choosing them carefully:
- Relevant: they closely mirror the real use case, not a toy version of it.
- Diverse: they cover edge cases and vary enough that Claude doesn't copy an accidental pattern, such as always using the same length.
- Structured: each one is wrapped in a tag named
example, with several together inside a tag namedexamples, so Claude can tell them apart from the instructions.
Start with one example. Add more only if the output still misses. A first prompt crowded with examples can lock in patterns you never intended, and refining from the first response is covered under prompt diagnosis and iteration.
Checkpoint 9 of 9· Match them up
A team lead's few-shot prompt for customer-reply drafts keeps going wrong. Match each problem to the guideline that fixes it.
Tap a term, then the definition that fits it.
Claude pays close attention to examples, including details you didn't mean to teach. Examples need to match the real use case. They need to vary so Claude doesn't copy an accidental pattern. They need to be wrapped in example tags so Claude can tell them apart from the instructions. Start with one and add more only if the output still misses.
“Diverse: Cover edge cases and vary enough that Claude doesn’t pick up unintended patterns.”Source: platform.claude.com
Sources
(7)
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“Think of Claude as a newly-hired contractor.”
↩︎ What Claude does not know when you start - 2.
“What generative AI knows comes entirely from training data and is frozen at the knowledge cutoff.”
↩︎ What Claude does not know when you start“What generative AI knows comes entirely from training data and is frozen at the knowledge cutoff.”
↩︎ Exam trap 1 - 3.https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practicesOfficial docs
“Show your prompt to a colleague with minimal context on the task and ask them to follow it.”
↩︎ What Claude does not know when you start“Even a single sentence makes a difference”
↩︎ The parts of a prompt and where they go“Place your long documents and inputs near the top of your prompt, above your query, instructions, and examples.”
↩︎ The parts of a prompt and where they go“Wrapping each type of content in its own tag (for example, <instructions>, <context>, <input>) reduces misinterpretation.”
↩︎ The parts of a prompt and where they go“When using multiple documents, wrap each document in <document> tags with <document_content> and <source> (and other metadata) subtags for clarity.”
↩︎ The parts of a prompt and where they go“Be specific about the output format and constraints you want.”
↩︎ Choosing an output format for the reader“Claude can generalize from the explanation.”
↩︎ Choosing an output format for the reader“A few well-crafted examples (known as few-shot or multishot prompting) improve accuracy and consistency.”
↩︎ Showing the format with examples“Relevant: Mirror your actual use case closely.”
↩︎ Showing the format with examples“Think of Claude as a brilliant but new employee who lacks context on your norms and workflows.”
↩︎ Key concept“Place your long documents and inputs near the top of your prompt, above your query, instructions, and examples.”
↩︎ Exam trap 2“Diverse: Cover edge cases and vary enough that Claude doesn’t pick up unintended patterns.”
↩︎ Checkpoint - 4.https://support.claude.com/en/articles/7996857-my-prompt-isn-t-giving-me-a-helpful-answerOfficial docs
“Pretend you are giving these instructions to someone with no background knowledge about what you are asking.”
↩︎ What Claude does not know when you start - 5.https://academy.claude.com/courses/ai-fluency-for-builders/description-building-great-thingsOfficial docs
“AI cannot hear what the user did not say.”
↩︎ Work from purpose toward wording“Every adjective should be a decision you can defend.”
↩︎ Work from purpose toward wording - 6.
“This exercise teaches you to prompt from intent, not just from instruction.”
↩︎ Work from purpose toward wording“The gap between what you say and what you mean is where most steerability failures live.”
↩︎ Exam trap 4“Then re-prompt with the goal stated explicitly alongside the instruction”
↩︎ Checkpoint - 7.https://claude.com/blog/best-practices-for-prompt-engineeringSecondary source
“Tell the AI what TO do instead of what NOT to do”
↩︎ Choosing an output format for the reader“Describing how the output will be used”
↩︎ Choosing an output format for the reader“Pro tip: Start with one example (one-shot).”
↩︎ Showing the format with examples“Describing how the output will be used”
↩︎ Exam trap 3“Instead of: "Do not use markdown in your response" Try: "Your response should be composed of smoothly flowing prose paragraphs"”
↩︎ Checkpoint