20 Best Claude Prompts, Built on Anthropic's Guidance

A prompt library is only useful if you know why the prompts work. So this one starts with what Anthropic actually tells people to do, then shows those techniques in prompts you can run today. You won't find "secret" phrases here, or claims about one model beating another.

What Anthropic says makes a Claude prompt work

Anthropic's prompting best practices compare Claude to a "brilliant but new employee who lacks context on your norms and workflows." The test it suggests is simple: if a colleague with little context would be confused by your prompt, Claude will be too. Here's what the docs recommend.

• Be clear about the output. Specify the format and constraints, and use numbered steps when order matters.
• Explain why. A reason behind an instruction lets Claude generalize from it.
• Show examples. The docs suggest 3 to 5, wrapped in example tags, relevant to your case and varied enough to cover edge cases.
• Use XML tags when a prompt mixes instructions, context, examples and inputs.
• Put long documents first, with your question near the end, and ask for relevant quotes before the answer.
• Say what to do, rather than only what to avoid.
• Allow uncertainty. Anthropic's guide to reducing hallucinations suggests giving Claude permission to say it doesn't know, and asking it to cite a quote for each claim.

Advice also changes by model, and that's the part older prompt lists miss. Several current models have their own prompting page. Asking for reasoning in thinking tags may be declined on Fable 5.1, Fable 5, Opus 5.5, Opus 5 and Sonnet 5.5, where the docs say to rely on built-in thinking. A closing "verify your answer" line usually helps, but on Opus 5 the docs say to remove it because it can cause over-verification. Our 12 prompt patterns cover the techniques that carry across assistants.

How to use these prompts in Claude

Most of these work in a plain chat. A few depend on features worth setting up first, all described in Anthropic's help center.

• Projects have their own chat history, knowledge base and instructions. All users get them, with a maximum of five on a free account. Chats don't share context unless the information is in the knowledge base.
• Artifacts open beside the conversation: documents, dashboards, small interactive tools. They're on all plans, and need code execution and file creation turned on in your capabilities settings.
• File creation covers Excel, PowerPoint, Word and PDF files, on all plans.
• Web search is turned on from the + menu in a chat, and Anthropic says that in its new Claude experience there is no toggle because Claude searches when it helps.
• Memory and chat search. Memory is on by default on Free, Pro, Max and Ultra. Searching past chats needs a paid plan.

Which prompt for which job

Writing and editing prompts

Adjectives like "witty" rarely get you your voice. Samples do. Writing marketing copy instead? We keep a separate set of AI prompts for marketing.

1. Write in my voice from real samples

Uses Anthropic's example tags. Give it 3 to 5 samples that differ in topic.

<instructions>
Write a [type of piece] about [topic] for [reader]. Length: [word count].
Match the voice in the examples: sentence length, word choice, and how the writer opens and closes. Do not reuse their phrases.
Use only the facts in my notes. Where a fact is missing, write [NEEDS FACT].
</instructions>
<examples>
<example>[sample 1]</example>
<example>[sample 2]</example>
<example>[sample 3]</example>
</examples>
<notes>[your raw points]</notes>

2. Edit without losing my meaning

Gives Claude the reason behind the edit, which Anthropic says helps it target the result.

<draft>[paste draft]</draft>

Edit this for [reader], who needs to [what they must do after reading]. They will read it on [phone, in a meeting, etc.], so the main point has to come first.
Keep my facts and my argument. Cut repetition, and replace vague phrases with the specific thing.
Return the edited version, then list any change that alters meaning so I can approve it.

3. Turn messy notes into a document

For call notes, voice memo transcripts and brain dumps.

<notes>[paste notes]</notes>

Turn these notes into a [memo, brief, proposal] for [reader].
Structure: the decision or ask first, then the background, then next steps with owners.
Write in full paragraphs, with a heading for each part.
Do not add anything that is not in the notes. Put open questions in a final section called Still unclear.

4. Standing project instructions

Paste this into a project's instructions, not a chat, so it applies to every chat in that project.

You are helping [team or person] with [ongoing work].
Reader: [who reads the output]. Voice: [3 plain descriptions, plus a pointer to the sample file in this project].
Default format: [e.g. short paragraphs with one heading per idea].
These documents reach [customers, the board], so tie each number to a file in this project.
If a fact is not in the project knowledge, say you do not have it. If a request could mean two things, ask one question before you start.

Research and long-document prompts

These prompts make the answer auditable. If you can't trace a sentence back to a quote or a link, don't rely on it.

5. Quote first, then answer

Anthropic's pattern for long documents: material on top, quotes extracted first, question last.

<documents>
<document index="1">
<source>[file name]</source>
<document_content>[paste text]</document_content>
</document>
<document index="2">
<source>[file name]</source>
<document_content>[paste text]</document_content>
</document>
</documents>

Find word-for-word quotes relevant to this question and put them in <quotes> tags, each with its document number.
Then answer in <answer> tags, using only those quotes.
If the documents do not answer it, say so. Question: [your question]

6. Pull out what a contract asks of me

Helps you prepare for a review. It does not replace legal advice.

<document>[paste contract or policy]</document>

I am the [customer, vendor, employee] in this document.
Make a table of what I must do: obligation, deadline, what happens if I miss it, and the exact clause.
Then list terms that look unusual to a non-lawyer, and 5 questions to ask a lawyer.
If you are not sure what a clause means, say so rather than guessing.

7. Compare two versions of a document

For redlines, updated policies and vendor terms that changed without notice.

<document index="1"><source>Old version</source><document_content>[paste]</document_content></document>
<document index="2"><source>New version</source><document_content>[paste]</document_content></document>

List every change that affects [what you care about, e.g. price, liability, timelines].
For each one, quote the old and new wording and explain the effect in one sentence.
Ignore formatting and wording changes that do not alter meaning, then state how many you ignored.

8. Run a web research brief

Use it with web search. Anthropic says responses that use search include citations, so open them.

Search the web to answer: [question].
A good answer: [what you need to decide, how recent the sources must be].
Prefer primary sources such as [official docs, filings, the vendor's own site]. Check each key claim against a second source where you can.

Return a table with claim, source link, source date and whether a second source confirmed it.
Then list where sources disagree and what you could not confirm.

Data and spreadsheet prompts

Anthropic says Claude can analyze uploaded data and build Excel models with working formulas. That doesn't make the numbers right, so open the file and test a row by hand.

9. Audit a spreadsheet before analysis

Attach the file first. Bad columns are cheaper to catch now than after the chart.

I have attached [file]. Do not analyze it yet.
Tell me what each column appears to contain, its type, and the share of blank values.
List data problems: duplicates, mixed formats, impossible values, totals hiding inside rows.
Propose a fix for each, then stop and wait for me to choose. Do not change the file until I answer.

10. Build an Excel model with working formulas

Relies on Claude's file creation. Check the formulas yourself before anyone uses the numbers.

Create an Excel file that models [what you are modeling] over [period].
Inputs, on their own sheet: [list each with its value and source].
Calculations: [how each output is worked out].
Use live formulas, not pasted values, so changing an input updates everything.
Add one sheet that lists every assumption, and flag any input I did not give you. Then explain how to test the model by hand with one example.

11. Classify feedback into a fixed format

Anthropic's consistency advice in practice: define the format and show examples.

Classify each item in <feedback> as JSON with these keys: "theme", "sentiment", "product_area", "quote".
Allowed themes: [list]. If none fits, use "other". The quote must be copied word for word.

<examples>
<example>[input and the JSON you want]</example>
<example>[a second, different case]</example>
<example>[an edge case]</example>
</examples>

<feedback>[paste items]</feedback>

Planning and decision prompts

Claude will happily plan with the little you gave it. These three make it ask, label and argue first.

12. Interview me before you write the plan

Useful when you know the goal but haven't worked out the details.

I want to [goal] by [date]. My constraints: [budget, people, hours].
Before you plan anything, ask me up to 5 questions, one at a time, about what would change the plan most. Wait for each answer.
Then write the plan as phases with a deadline, an owner and a way to tell the phase is done.
Name the assumption most likely to be wrong.

13. Write a decision memo

Makes the tradeoff visible before you commit.

I need to choose between [option A] and [option B].
What matters most, in order: [criteria]. What I know so far: [facts]. What I am afraid of: [worry].

Write a one-page memo: your recommendation first, then the best case for each option, what each one costs if it goes wrong, and which is easier to reverse.
Mark each statement as FACT (from my notes) or JUDGMENT (yours). Then name the one piece of information that would change your recommendation.

14. Run a pre-mortem

Most useful right before a launch, a hire or a big purchase.

<plan>[paste plan]</plan>

Imagine it is [date] and this plan has failed. Write the 5 most likely reasons, specific to this plan rather than generic project risks.
For each, give the early warning sign, who would see it first, and the cheapest thing we could do this week to reduce the risk.
Rank them by how much damage they would do.

Artifact and file prompts

Say "create an artifact" or name the file type, because the output format is part of the instruction.

15. Build a small calculator as an artifact

Anthropic lists small interactive tools among the things Claude can make as artifacts.

Create an artifact: an interactive calculator for [what it works out, e.g. monthly cost by seat count].
Inputs: [each input, with its default and allowed range].
Formula: [write it out].
Show the result as it changes, plus a one-line explanation of how it was calculated.
Make it usable on a phone. Label any default that is my assumption.

16. Turn a data file into a dashboard

Check the data audit in prompt 9 first, then let someone click through the result.

I have attached [file]. Create an artifact: a one-page dashboard for [reader], who needs to answer [question].
Show [3 to 5 metrics], with [the breakdown that matters] as a filter.
Above the charts, write 3 sentences on what stands out, each one backed by a number on the page.
List any rows you excluded and why.

17. Make a deck from a document

Uses Claude's PowerPoint file creation. Read every slide before you present it.

<document>[paste or attach the source]</document>

Create a PowerPoint file of [number] slides for [audience]. They need to decide [decision] by the end.
Each title is a full-sentence takeaway. Use at most 3 short bullets per slide, and put the detail in speaker notes.
Use only facts from the document. Add a final slide listing the source section behind each number.

Coding prompts

These work in chat if you paste the code, or in Claude Code where it can read your files. Choosing a tool? See Cursor vs Claude Code, or the wider Cursor, Replit and Claude comparison. For bigger builds, prompt stacking breaks the work into ordered steps, and these prompting techniques for code quality go deeper.

18. Debug without guessing

Anthropic's docs suggest having Claude check the code before it answers questions about it.

Bug: [what happens]. Expected: [what should happen]. Steps to reproduce: [steps]. Error output: [paste].

Start by listing the files you opened. Treat anything you have not seen as unknown, and label it that way.
Give me your top 3 likely causes, with the file and line behind each, and how to confirm each one.
Fix the confirmed cause only, then tell me how to check the fix.

19. Make the smallest correct change

Anthropic flags over-engineering on some models and says to ask for minimal solutions.

Change requested: [change].
Before editing, name the fewest files and lines this can be done in. Leave everything else as it is, and put anything else you would improve under a heading called Seen but not changed.
Check the result against realistic inputs such as [example inputs], not only what the test suite covers. If a test seems to expect the wrong result, stop and show me why.
When you finish, say why each file you edited had to change.

20. Review a diff like a careful teammate

A first pass before human review, not a replacement for it.

<context>This change is meant to [goal]. It runs in [environment].</context>
<diff>[paste diff]</diff>

Review for bugs, missing edge cases and security problems. Skip style preferences.
For each finding, give the file and line, what breaks, the input that triggers it, and a suggested fix.
Rank by severity. If you find nothing serious, say so and do not pad the list.

Common mistakes with Claude prompts

• Rules with no reason. Anthropic's own example: "never use ellipses" works better when you say a text-to-speech engine will read the output.
• Copying old prompts forward. Thinking tags and verification lines can backfire on some current models. Check Anthropic's guidance for your model.
• Examples that are all alike. Claude can pick up patterns you didn't intend, so vary them.
• Expecting chats in a project to know each other. Move what matters into the knowledge base.
• Trusting output unread. Anthropic says models can generate factually incorrect text, so open the citations and recheck the math.

Frequently Asked Questions

What are the best Claude prompts?
The best Claude prompts state the output you want, explain why it matters, separate instructions from inputs with XML tags and include your real material. The 20 prompts on this page follow that structure for writing, research, data, planning, artifacts and code.

Do I need XML tags in Claude prompts?
No, a short request works without them. Anthropic recommends XML tags when a prompt mixes instructions, context, examples and variable inputs, because they reduce misinterpretation. Any consistent, descriptive tag names will do.

Should I give Claude a role?
It can help. Anthropic's API docs set a role in the system prompt and say even one sentence makes a difference. In the Claude apps, its help center gives answering from the perspective of a specific role or industry as an example of what project instructions can do.

Do these prompts work on the free Claude plan?
Most do. Anthropic's help center lists projects (up to five on a free account), artifacts and file creation for free users, while searching past chats needs a paid plan. Anthropic also warns free users that a long document fetched from a link can use a significant share of the usage limit.

Where does Anthropic publish its prompting guidance?
In the Claude Platform Docs. The Prompting best practices page covers techniques for current models and links to separate prompting pages for individual models. Anthropic also has an interactive tutorial on GitHub, written for Claude 3, and says to follow the best practices page where the two differ.