Guide · Claude
The Best Claude Prompts
A curated, copy-ready library of the highest-leverage Claude prompts — engineered for Claude's strengths in long-context reasoning, careful writing and structured analysis.
PromptInFlow AI Guides
General AI Models
Image & Video Models
AI Task Guides
- AI WritingRead guide
- AI MarketingRead guide
- AI CodingComing soon
- AI BusinessComing soon
- AI EducationComing soon
Claude rewards a different kind of prompting than ChatGPT. Where most LLMs default to confident, polished surface-level answers, Claude reasons carefully, asks fewer rhetorical questions, and almost always benefits from explicit structure — XML tags, numbered constraints, a clearly stated role. Treat it like a meticulous senior teammate that prefers a one-page brief over a one-line ask, and the output quality jumps from "good enough" to "shipped." The prompts in this guide are written specifically for how Claude thinks: they front-load context, separate instructions from data, and ask Claude to plan before it writes whenever the task involves reasoning.
This is a working reference, not a roundup. Every prompt has been tested on the current Claude lineup (Claude 3.5 Sonnet, Claude 3.5 Haiku, and Claude 4 Opus / Sonnet via Anthropic's API and Claude.ai), grouped by the job you're trying to get done. Expect prompts for long-form writing and editing, document analysis over 100k-token contexts, research synthesis, code review and refactoring, structured business analysis, and the careful, nuanced reasoning Claude is genuinely best at. You'll also find the prompt-engineering patterns Anthropic's own research surfaces — XML tagging, system-prompt role priming, prefilling — applied to real workflows.
Who is this guide for? Writers and editors who want a model that respects voice and length constraints. Researchers and analysts pasting in large PDFs, transcripts or codebases and asking for honest, hedged conclusions. Engineers using Claude as a code-review partner that flags edge cases other models miss. Founders and operators who'd rather have a model push back on a bad idea than agree with it. If you've used Claude and thought "this feels different from ChatGPT — I'm not getting the most out of it," this guide closes that gap.
To get the best from Claude, treat every prompt as a structured brief. Wrap inputs (documents, code, examples) in XML tags like <document> or <examples> so Claude can distinguish data from instructions. Ask Claude to think before answering for any non-trivial task ("Before responding, plan your approach inside <thinking> tags"). Give it room to say it doesn't know — Claude is unusually honest about uncertainty when you invite it. Bookmark this page; we refresh it whenever Anthropic ships a new model or behavior change.
Editorial standards
How we selected these prompts
Why each prompt earns its place in the library.
Every prompt in this guide earns its place. We test against Claude's current production lineup (Sonnet 3.5, Haiku 3.5 and Opus / Sonnet 4) and only ship prompts that produce a usable first output on a fresh conversation — no chain of follow-ups required. Prompts come from our own daily workflows, patterns Anthropic publishes in their prompting documentation, and refinements contributed by the Prompt InFlow community.
Built for Claude's strengths
Every prompt leans into what Claude does best — long-context reasoning, careful analysis, on-tone writing and honest hedging — instead of treating Claude like a generic chatbot.
Structured by default
We use XML tags, explicit roles and numbered constraints because Claude responds measurably better to structured input. Each prompt models the pattern so you can reuse it.
Adaptable placeholders
Every template uses clear {placeholders} so the same prompt adapts from a solo consultant to an enterprise team by swapping five words, not rewriting from scratch.
No filler
We'd rather publish 40 prompts you'll reuse every week than 400 you'll scroll past. If two prompts overlap, the weaker one is cut.
Who this guide is for
This guide is built for anyone who wants Claude to do serious thinking work — not just rephrase a paragraph.
Long-form writers
Essayists, ghostwriters and content leads who need on-voice drafts that don't drift after 1,500 words.
Researchers & analysts
People pasting 100k-token documents, transcripts or datasets and asking for synthesis they can actually trust.
Engineers
Using Claude for code review, refactors and architectural sketches where edge cases and honest pushback matter.
Founders & operators
Drafting strategy memos, board updates and SOPs — and asking Claude to steelman the counter-argument.
Legal, policy & compliance
Reviewing contracts, policies and regulatory text where careful, hedged language beats confident guessing.
Educators & students
Building explainers, study plans and Socratic dialogues that go deeper than a one-shot summary.
In this guide
What you'll learn
Skim the highlights, then jump to the section you need.
How Claude is different from ChatGPT
Where Claude wins (long context, careful reasoning, on-tone writing) and where it doesn't — so you pick the right tool for the job.
40+ copy-ready Claude prompts
Writing, analysis, research, coding and business prompts you can paste into Claude.ai or the API and run today.
Claude-native techniques
XML tagging, system-prompt role priming, prefilling, and the "think before answering" pattern that lifts reasoning quality.
How to scale prompts to 100k+ tokens
Patterns for putting documents at the top, instructions at the bottom, and asking Claude to cite passages it relied on.
Framework
How great Claude prompts work
The six-part anatomy behind every prompt that consistently ships usable output.
Claude's prompting style is closer to a structured brief than a conversation. Anthropic's own research repeatedly shows the same handful of moves lifting output quality — and they map cleanly onto a framework we use across the library: R-S-C-I-O-T (Role, Structure, Context, Instructions, Output, Thinking). Use all six on complex tasks; use the first four on everything else.
Role
Open with a clear role in the system prompt or first message. "You are a senior policy analyst reviewing draft legislation" calibrates Claude's vocabulary, tone and rigor before it sees the task — and Claude respects role framing more consistently than most models.
Structure (XML tags)
Wrap inputs in XML tags so Claude can tell data from instructions: <document>, <examples>, <criteria>, <transcript>. Tagged input measurably reduces drift, especially on long contexts, and lets you reference parts back to Claude ("using the rubric in <criteria>…").
Context
Give Claude the background a new contractor would need: who the audience is, what's been tried, what the success criteria look like, what's off-limits. Claude under-performs on under-specified tasks — it would rather have too much context than guess.
Instructions
Number multi-step tasks (1, 2, 3) and put the most important constraints last — Claude weights end-of-prompt instructions slightly higher. State explicitly when Claude is allowed to say "I don't know" or "the document doesn't cover this."
Output format
Spell out the shape of the answer: a memo with H2s, a JSON object with these keys, a markdown table, a bulleted list grouped by severity. For structured output, prefilling the assistant turn with the opening bracket or heading dramatically improves reliability.
Thinking (plan-before-answer)
For any non-trivial task, instruct Claude to plan inside <thinking> tags before writing the final answer. This single move is the biggest quality lever for analysis, code review, and multi-step reasoning — and you can ask Claude to discard the thinking from the final output.
Three examples — and why they work
Long-document analysis
You are a senior research analyst. I will give you a long document. Your job is to extract the three strongest arguments, the three weakest, and any factual claims that need verification.
<document>
{paste full document here}
</document>
Instructions:
1. Before answering, plan your approach inside <thinking> tags.
2. For every claim, cite the exact passage in quotes.
3. If the document does not support a conclusion, say so explicitly.
Output: a markdown report with three H2 sections (Strongest arguments, Weakest arguments, Claims to verify), each with quoted evidence.On-voice long-form writing
You are a ghostwriter writing in the voice of {author name}. Below are two short samples of their published work.
<voice_samples>
<sample>{paste 200-word sample 1}</sample>
<sample>{paste 200-word sample 2}</sample>
</voice_samples>
Task: draft a 900-word essay on {topic}. Match cadence, sentence length distribution and vocabulary range. Avoid: em-dashes used as drama, "in conclusion," any sentence opening with "Moreover."
Output: clean markdown, no preamble. Title + 4–6 sub-headed sections.Code review with severity grouping
You are a staff engineer doing a PR review on a TypeScript codebase. Review the function below for correctness, edge cases, readability and performance — in that priority order.
<code>
{paste function}
</code>
Before responding, plan inside <thinking> tags: list potential edge cases, then walk through the function against each.
Output: a markdown list grouped by severity (Blocker, Nit, Optional). For every item: cite line numbers, suggest a concrete fix, and explain the failure mode. If the code looks correct, say so plainly instead of inventing nits.Browse by use case
Featured prompt categories
Jump into the library that matches what you're working on — each category includes guidance on when to use it and what good output looks like.
- When to use
- Reach for Claude on long-form writing — essays, reports, ghostwritten posts, white papers, scripts — and any rewrite where matching a specific voice matters more than speed.
- What you can achieve
- On-voice drafts that don't drift after 1,500 words, careful editing that preserves the writer's intent, and rewrites grounded in two or three pasted reference samples.
- Best practices
- Paste 1–2 real writing samples inside <voice_samples> XML — Claude calibrates voice from examples more reliably than from "write like" instructions.
- Specify length in words AND structure (sections, sub-headers) — Claude respects both constraints precisely.
- Ban the AI-tells in your constraints ("in today's fast-paced world," "navigating the landscape," rhetorical questions) — Claude obeys explicit bans well.
- When to use
- Use Claude for strategy memos, competitive analysis, board updates, financial commentary and SOPs — anything where you'd rather have honest pushback than confident agreement.
- What you can achieve
- Decision-ready memos with explicit trade-offs, steel-manned counter-arguments, sanity-checked financial narratives, and SOPs that survive a real team using them.
- Best practices
- Ask Claude to steelman the opposing position before recommending — it surfaces blind spots more honestly than most models.
- Give Claude permission to say "I don't have enough information to answer that" — it will use it, and the output gets sharper.
- Request decision-ready format (Recommendation → Rationale → Trade-offs → Open questions), not exhaustive analysis.
- When to use
- Use Claude for code review, refactors, architectural sketches, test generation and any task where catching edge cases matters more than raw speed.
- What you can achieve
- Pull-request-quality reviews with cited line numbers, regression-safe refactors with tests, and architecture trade-off analyses you can drop into a design doc.
- Best practices
- Wrap code in <code> tags and ask Claude to enumerate edge cases in <thinking> before judging — this is the single biggest lever on review quality.
- State runtime, language version and constraints up front ("Node 20, TypeScript strict, no new dependencies").
- For non-trivial fixes, ask for tests alongside the fix and reasoning for each test case.
- When to use
- Use Claude for explainers, study plans, Socratic dialogues, lesson outlines and any teaching task where careful, hedged, level-appropriate language matters.
- What you can achieve
- Explainers at the exact reading level you specify, Socratic question chains that build understanding instead of dumping answers, and feedback that names the misconception, not just the wrong answer.
- Best practices
- Specify reader level explicitly ("explain to a curious 12-year-old," "to a first-year CS student") — Claude calibrates precisely to age and background.
- Ask for a Socratic dialogue (questions only) when the goal is understanding, not recall.
- When grading, ask Claude to identify the underlying misconception, not just mark right/wrong.
Hand-picked
Curated prompts
The prompts we reach for most often — copy, paste, adapt.
Senior Engineer Code Review
Get a staff-engineer-grade review of a diff with risk callouts, simpler alternatives and test gaps.
You are a staff engineer. Review this diff and return: (1) correctness risks, (2) simpler alternatives, (3) missing tests, (4) one nit. Be specific. Diff: {diff}
High-Converting Cold Sales Email
Write a 4-line cold outreach email with a personalized hook, value prop, social proof and soft CTA.
You are a B2B sales copywriter. Write a 4-sentence cold email to {persona} at {company}. Open with a specific observation about {trigger_event}, then connect it to {value_prop}. Include one short social proof line and a low-friction CTA.
SEO Topic Cluster Builder
Generate a pillar page plus 12 supporting article ideas with target keywords and search intent.
Act as an SEO strategist. For the seed topic {topic}, produce one pillar page idea and 12 supporting articles. For each: working title, primary keyword, search intent, and one internal link target.
At a glance
Quick reference tables
Cheat-sheets for picking the right model and the right prompt style.
Claude models at a glance
Pick the model that matches the job — Claude's tiers differ more than ChatGPT's.
| Model | Best for | Watch out for |
|---|---|---|
| Claude 4 Opus | The hardest reasoning, research, and code tasks | Slowest and most expensive — overkill for routine writing |
| Claude 4 Sonnet | Daily long-form writing, analysis, coding | Slightly slower than 3.5 Sonnet for short tasks |
| Claude 3.5 Sonnet | Balanced speed/quality default for most workflows | Older — Claude 4 Sonnet now beats it on most benchmarks |
| Claude 3.5 Haiku | High-volume, latency-sensitive tasks (classification, summaries) | Weaker at multi-step reasoning and nuanced writing |
Claude vs ChatGPT — pick the right tool
Both are excellent. They're best at different things.
| Task | Pick Claude when | Pick ChatGPT when |
|---|---|---|
| Long-form writing | Voice/tone matters and length is 1,000+ words | You need a fast first draft under 500 words |
| Document analysis | Document is 30k+ tokens and citations matter | Document is short and you want quick takeaways |
| Code review | Edge cases, correctness, honest pushback | Boilerplate generation, quick fixes, scaffolding |
| Strategy / decisions | You want a model that steel-mans the counter-argument | You want fast brainstorming and many variants |
| Multimodal (image / voice) | Image analysis with careful captioning | Anything voice, real-time, or image generation |
Claude-specific prompting techniques
Patterns that lift quality on Claude specifically.
| Technique | When to use | Example trigger |
|---|---|---|
| XML tagging | Any prompt with pasted inputs (docs, code, examples) | "<document>…</document>" and "<criteria>…</criteria>" |
| Plan-before-answer | Reasoning, analysis, code review, multi-step tasks | "Before responding, plan inside <thinking> tags." |
| Prefilling | Strict output formats (JSON, markdown tables, memos) | Start the assistant turn with "{" or "# Memo" |
| Permission-to-not-know | Anything where hallucination is costly | "If the document doesn't cover this, say so explicitly." |
Method
How to write better prompts
Four habits that lift the quality of every prompt you write.
Wrap your inputs in XML
Any time you paste a document, code, examples or a transcript, wrap it in a tag like <document> or <code>. Claude separates data from instructions far more reliably with XML than with markdown fences — and it lets you reference the input back later in the prompt.
Ask Claude to plan before answering
For any reasoning, analysis or code task, instruct Claude to plan inside <thinking> tags before writing the final answer. It's the single biggest quality lever — and you can ask Claude to omit the thinking block from the final output.
Give explicit permission to say "I don't know"
Claude is unusually honest about uncertainty when you invite it. "If the document doesn't support a conclusion, say so explicitly" prevents the polite hallucinations that creep in when the model thinks it has to answer.
Prefill the format
For strict outputs (JSON, tables, memos), start the assistant turn with the opening bracket, header, or first row. Prefilling dramatically improves format adherence — and is uniquely powerful on Claude.
Avoid
Common mistakes
The patterns that quietly tank output quality.
- Pasting documents inline without XML tags — Claude blurs the line between data and instructions and the output drifts.
- Skipping the plan-before-answer step on reasoning tasks — Claude can do the reasoning, but it often won't unless you ask.
- Treating Claude like ChatGPT and asking for 10 punchy variants — Claude is better at one careful answer than many shallow ones.
- Putting the most important instructions at the top — Claude weights end-of-prompt constraints slightly higher.
- Forgetting to give Claude permission to refuse or hedge — without it, the model invents an answer to seem helpful.
- Using Claude for image generation, voice, or real-time tasks — those aren't its strengths; pick the right tool.
FAQ
Frequently asked questions
Quick answers about this guide and the prompts in it.
Keep reading
Related guides
The Best ChatGPT Prompts
A curated, copy-ready library of the highest-leverage ChatGPT prompts — organized by use case and tested across GPT-4o and GPT-5.
The Best Gemini Prompts
A curated, copy-ready library of the highest-leverage Google Gemini prompts — built around Gemini's strengths in multimodal reasoning, massive context and live Google data.
The Best AI Writing Prompts
Frameworks, model picks and copy-ready prompts for every writing task — from blog posts and emails to SEO content, product copy and storytelling.
The Best AI Marketing Prompts
A curated, model-aware library of marketing prompts that ship — built for the channels and tasks marketers actually run every week.
Explore
Related categories
Compatible models
Works well with
Fresh
Latest prompts
Recently added to the Prompt InFlow library.
Senior Engineer Code Review
Get a staff-engineer-grade review of a diff with risk callouts, simpler alternatives and test gaps.
You are a staff engineer. Review this diff and return: (1) correctness risks, (2) simpler alternatives, (3) missing tests, (4) one nit. Be specific. Diff: {diff}
High-Converting Cold Sales Email
Write a 4-line cold outreach email with a personalized hook, value prop, social proof and soft CTA.
You are a B2B sales copywriter. Write a 4-sentence cold email to {persona} at {company}. Open with a specific observation about {trigger_event}, then connect it to {value_prop}. Include one short social proof line and a low-friction CTA.
SEO Topic Cluster Builder
Generate a pillar page plus 12 supporting article ideas with target keywords and search intent.
Act as an SEO strategist. For the seed topic {topic}, produce one pillar page idea and 12 supporting articles. For each: working title, primary keyword, search intent, and one internal link target.
Veo Character Consistency Sheet
Keep the same character across multiple Veo shots with a reusable description block and shot list.
Character: 32-year-old woman, shoulder-length auburn hair, freckles, olive linen jacket. Reuse this exact description for every shot. Shot 1: medium close-up, golden hour. Shot 2: walking shot, overcast street. Shot 3: cafe interior, warm tungsten.
In summary
Conclusion
The ideas to carry forward — and where to go next.
If there's one idea to carry out of this guide, it's that Claude rewards structure. A clear role, XML-tagged inputs, numbered instructions, a plan-before-answer step, and explicit permission to say "I don't know" — those five moves separate prompts that produce shipped work from prompts that produce filler. None of them are clever; they're just respected by Claude more consistently than by any other model.
Treat every prompt in this library as a starting point. Copy it, replace the {placeholders} with something specific to your audience, paste your real inputs inside the XML tags, and iterate with short follow-ups rather than rewriting from scratch. Claude is unusually good at refinement — "tighten the second section, cut by 30%, keep the citations" almost always works on the first try.
Prompting is a skill, and it compounds. The more prompts you ship and refine on Claude, the faster you'll recognize when to reach for it instead of ChatGPT, Gemini or a generator — and when to switch. When you're ready for more, explore our other model guides, browse the prompt library by use case, or jump into our generators to build a Claude-ready prompt in seconds.
Continue Learning
Prompt Engineering Learning Path
A guided reading order through every Prompt InFlow flagship guide — start with the model you use most, then layer in topical skills.
- 1
Best ChatGPT Prompts
50+ tested ChatGPT prompts and the R-G-C-C-O-E framework that makes them ship usable output.
Model guide - 2You are here
Best Claude Prompts
Long-context reasoning, on-voice writing and XML-tagged briefs for Claude 3.5 and Claude 4.
Model guide - 3
Best Gemini Prompts
Grounded, multimodal and long-context prompts that lean into Gemini's real-time strengths.
Model guide - 4
Best AI Writing Prompts
Voice, structure and editing — the writing playbook across ChatGPT, Claude and Gemini.
Topical guide - 5
Best AI Marketing Prompts
SEO, content, social, email, ads and brand — model-matched prompts for every channel.
Topical guide
Build prompts faster with Prompt InFlow
Browse 500+ tested prompts across every major AI model — or use our generators to build your own in seconds.