Claude Prompt
Generator.
Build XML-structured prompts engineered for Claude. Extended thinking, few-shot slots, agent scaffolding — all wired up by default.
Works in Claude.ai, Claude Projects, and the Anthropic API. Free. No sign-up.
<role> You are a Senior engineer with 15+ years experience. </role> <task> [describe the task] </task> <method> Read the provided material carefully before responding. Quote the specific passages your analysis relies on inside <evidence> tags. Separate observation from inference — facts go in <facts>, your interpretation goes in <analysis>. Before your final answer, output 3–5 bullets of reasoning inside <thinking> tags, then the answer. </method> <output_format> - Format: XML-tagged sections - Wrap distinct sections in descriptive XML tags so downstream parsing is trivial. - If any required input is missing, ask ONE clarifying question inside <clarify> tags before proceeding. </output_format> <quality> - Ground every claim in the provided material or in your training data — and clearly distinguish between the two. - If you are uncertain, say "I'm not sure" rather than guessing. - Verify your output against the constraints before delivering. </quality>
Paste into claude.ai, a Projects system prompt, or the Anthropic API. Everything happens in your browser.
Prompt anatomy
Why Claude prompts look different.
Claude was trained to respect tag boundaries. Wrapping task, context, examples, and output in XML tags measurably improves instruction-following vs. markdown prose.
Giving Claude room to work — inside <thinking> tags or an extended-thinking turn — reliably lifts accuracy on reasoning-heavy tasks. The tokens are cheap; the accuracy gain isn't.
Claude pattern-matches on examples more strongly than on abstract rules. One concrete example beats three paragraphs of instructions whenever output shape matters.
Best practices baked in
What makes a prompt Claude-native.
- •Structure wins. Named tags (<task>, <context>, <examples>, <output_format>, <quality>) give Claude clear boundaries and let you extract specific sections from the response.
- •Evidence separates from inference. Ask Claude to quote the specific passage supporting each claim in <evidence> tags. Paraphrasing hides hallucinations; verbatim quoting exposes them.
- •Tell Claude what "done" looks like. Explicit output formats — schema, word count, tone — are followed more reliably than implicit ones. If you want JSON, provide the exact schema. If you want HIGH / MEDIUM / LOW risk labels, specify the options.
- •Give permission to ask. A built-in <clarify> escape valve stops Claude from inventing missing information. If a required input is absent, the model asks one question instead of bluffing.
Related tools
Keep going with Claude.
FAQ
Questions about Claude prompts.
Why does this generator use XML tags instead of markdown?+
Claude models were trained on a diet of XML-heavy data and demonstrably follow instructions more reliably when sections are wrapped in tags like <task>, <context>, and <output_format>. XML also gives you parseable structure in the response — if you ask for <plan> and <code>, you can extract each section programmatically. Markdown headings work, but XML is the native dialect.
What is extended thinking and when should I use it?+
Extended thinking is a feature available in Claude Opus 4.6 and Sonnet 4.6 where the model works through the problem in a <thinking> block before answering. Use it for complex reasoning — legal analysis, debugging, strategy, multi-step math. Skip it for simple tasks like summaries or rewrites where the extra tokens add cost without improving quality.
Does this work with Claude in Projects, Claude API, and claude.ai?+
Yes. The generated prompt is plain text with XML structure — it pastes into any Claude surface without modification. For Projects, drop the ROLE + METHOD + OUTPUT FORMAT + QUALITY sections into the system prompt and reserve the TASK section for each conversation. For the API, send the whole thing as a user message or split between system and user per your app.
Should I include examples? How many?+
Yes when the task format is specific or the success criteria are subtle. Claude's accuracy on structured outputs jumps sharply with 1–3 well-chosen examples. More than 5 tends to narrow the model's creativity. For tasks where the output is obvious (summarize, translate), examples aren't necessary.
What's the difference between this and the ChatGPT prompt generator?+
Claude responds best to XML tags, explicit thinking directives, and strong persona framing. ChatGPT does fine with markdown, prefers numbered lists, and often wants a 'role' statement. Our ChatGPT generator uses the RGC (Role-Goal-Context) framework; this one uses XML-structured, Anthropic-style prompting. The generated prompts are not interchangeable without modification.
Does it handle agent / tool-use prompts?+
Yes. The 'Agent / tool use' mode produces prompts that instruct Claude to plan before acting, wrap tool calls in the expected format, and update its plan after each observation. It pairs well with Claude's tool-use API and the Claude Agent SDK.
Are my inputs sent anywhere?+
No. The generator runs entirely client-side. Your task, context, and examples never leave your browser. We don't log, store, or transmit any input.