Keep this open while you work in Claude. It has the prompt formula that makes answers sharper, the XML tags and output modifiers Claude responds to, and the phrases that switch on Artifacts, web search, and thinking. Every table is copy-paste ready. For finished prompts built on these, see the best Claude prompts and Claude prompt templates.
The RTCF formula
Almost every strong Claude prompt has four parts: Role, Task, Context, Format.
| Part | What to write |
|---|---|
| Role | Who Claude should be — sets standards, vocabulary, and point of view. |
| Task | The one thing to do, stated precisely. |
| Context | Background, material, and constraints — wrapped in XML tags. |
| Format | The shape of the answer: length, structure, table/JSON/diff/Artifact. |
Act as a senior product marketer. [Role]
Write a launch announcement for our new analytics dashboard. [Task]
<context>
Audience: existing Pro-plan customers. New: real-time charts + CSV export. Brand voice: friendly, plain-spoken.
</context>
Under 150 words, one call to action ("View your dashboard"), no hype words. Return the subject line, then the email body. [Format]You don't need the literal labels once it's habit — just keep the four ideas present and in order.
XML tags — Claude's superpower
Claude was trained to respect XML tags. Wrap each input in a named tag so Claude knows what it is and never confuses your material with your instructions.
| Tag | Use it for |
|---|---|
| <document> … </document> | A report, contract, or PDF you pasted |
| <code> … </code> | Source code to review, fix, or explain |
| <error> … </error> | A stack trace or error message |
| <data> … </data> | A CSV, table, or dataset |
| <transcript> … </transcript> | A meeting or interview transcript |
| <samples> … </samples> | Examples of your voice or on-brand copy to match |
| <example> … </example> | A worked example of the output you want (multishot) |
Then reference them by name: "rewrite the text in <draft> using the voice in <samples>".
Role modifiers
Open with a role to set Claude's perspective and standards. In the API or a Project, put it in the system prompt so it persists.
| Role phrase | What it changes | Use for |
|---|---|---|
| Act as a staff software engineer | Idiomatic code, edge cases, trade-offs | Code review, architecture |
| Act as a strict line editor | Tightens, cuts, flags weak lines | Rewrites, proofreading |
| Act as a skeptical reviewer | Pushes back, lists risks and gaps | Plans, proposals, decisions |
| Act as a data analyst | Structured findings, caveats on the data | Spreadsheets, reports |
| Act as a patient tutor | Step-by-step, checks understanding | Studying, explanations |
| Act as my chief of staff | Prioritizes, organizes, decides | Planning, triage |
| Act as a subject-matter expert in [field] | Domain terms, current best practice | Technical or niche topics |
Output-format modifiers
Claude follows explicit format instructions reliably. Add one to the Format part of your prompt.
| Say this | You get |
|---|---|
| Return a table with columns X, Y, Z | A clean markdown table, no preamble |
| Return valid JSON matching this schema: {...} | Parseable JSON only, no prose |
| Point to the specific lines; don't rewrite the file | A scoped change, not a full rewrite |
| Put it in an Artifact | A live, editable panel you can iterate on |
| Give me numbered step-by-step instructions | Ordered steps to follow |
| Keep it under [N] words | A hard length cap Claude respects |
| Answer first, then explain | The conclusion up front, reasoning after |
Thinking
Extended thinking makes Claude reason step by step before answering. Match it to the difficulty.
| Situation | Thinking? |
|---|---|
| Quick rewrite, lookup, simple summary, tone change | Off — keep it fast |
| Everyday drafting, analysis, multi-step instructions | Optional |
| Debugging, code architecture, math/logic, forecasting | On — "think step by step before answering" |
| High-stakes decisions where a wrong answer is costly | On |
On Fable 5, adaptive thinking is always on and scales itself, so you rarely toggle anything.
Feature triggers
These phrases and features change what Claude can do. Include one when you want live data, a live document, or persistent context.
| Phrase / feature | What it does |
|---|---|
| "Put this in an Artifact", "make it editable" | Renders code, docs, or a working app in a live side panel |
| "Search the web", "cite current sources" | Web search — live results with citations |
| "Use the analysis tool", "compute this with code" | Runs code on your data for exact results and charts |
| "Think step by step before answering" | Turns on visible reasoning before the answer |
| Projects | Saved workspace with shared files, style, and instructions |
| Styles | A saved writing voice applied to every reply |
| Claude Code | Agentic coding in your terminal/IDE across a real repo |
Model picker
Same prompt text; pick the model for the job.
| Model | Best for |
|---|---|
| Claude Opus 4.8 | Strong default — coding, hard reasoning, long tasks |
| Claude Fable 5 (Mythos-class) | The most complex, long, multi-step work; 1M context, always-on thinking |
| Claude Sonnet 5 | Agentic all-rounder, great value for everyday and browser/terminal work |
| Claude Haiku 4.5 | Fast and cheap for high-volume simple tasks |
Example prompts
Six complete prompts assembled from the tables — a role, a format, and where useful a feature trigger, wrapped in RTCF.
1. Grounded market summary
Act as a market analyst. Give me the current state of the [market] market. Search the web and cite current sources. Return a table (Brand, Flagship, Price band, Standout feature), then 3 bullets on where it's heading. Under 300 words.Best for: fast, sourced overviews. (Web search on)
2. Code review
Act as a staff engineer. Think step by step. Review this for correctness, edge cases, and security.
<code>
[paste code]
</code>
List issues most severe first with the line and a fix. Don't rewrite the whole file.Best for: catching real bugs. (Opus 4.8, thinking on)
3. Doc summary
Summarize this for a [role] who cares about [priority].
<document>
[paste or attach the document]
</document>
3-sentence executive summary, then 7 key points in priority order, then any numbers and deadlines verbatim.Why it works: the tag plus the priority focus keeps the summary useful, not generic.
4. Rewrite in my voice
Learn my voice from <samples> and write [the piece] in it.
<samples>
[paste 2-3 pieces of your writing]
</samples>
First describe my voice in 5 bullets, then write the piece. Don't drift into generic style.Best for: on-voice drafts. (Save as a Style to reuse)
5. Data question
Answer this from my data. Use the analysis tool; show your work.
<data>
[attach the CSV]
</data>
Question: [precise question]. Give the exact number, the calculation, and any assumption.Best for: trustworthy numbers. (Analysis tool on)
6. Build a tool
Build a working single-file [tool] as an Artifact. One self-contained HTML file, no dependencies, clean responsive UI, handle empty and error states. After it works, suggest 3 features to add.Best for: instant prototypes. (Artifact)
Want ready-made skeletons? Grab Claude prompt templates, or browse the best Claude prompts for finished examples.
Frequently Asked Questions
What is the RTCF prompt formula for Claude?
RTCF stands for Role, Task, Context, Format. Give Claude a role to set standards and vocabulary, state the task precisely, provide the context it needs (wrapped in XML tags), and specify the output format. Keeping all four present removes most of the guesswork that causes vague or off-target answers.
Why does Claude respond so well to XML tags?
Claude was trained to pay attention to XML tags, so wrapping each input in a named tag — <document>, <code>, <data>, <example> — tells Claude exactly what each block is and where it starts and ends. This keeps long or multi-part prompts organized and stops Claude from confusing your source material with your instructions. Tags are optional but reliably improve control.
When should I turn on extended thinking?
Turn it on for hard reasoning — math and logic, code architecture and debugging, multi-document synthesis, and high-stakes decisions — where you want Claude to reason before answering. Leave it off for quick rewrites, simple lookups, and formatting where speed matters. Fable 5 runs adaptive thinking automatically and scales it to the task.
Which Claude model should I pick?
Opus 4.8 is the strong default for coding and hard reasoning. Fable 5, the top Mythos-class model, is best for the most complex, long tasks with its 1M-token context and always-on adaptive thinking. Sonnet 5 is the agentic all-rounder and great value for everyday and browser/terminal work. Haiku 4.5 is the fast, cheap option for high-volume simple tasks.
Do format instructions like "return only a diff" work?
Yes. Claude follows explicit output-format instructions reliably — ask for a diff, a table, JSON, numbered steps, an Artifact, or a strict word count and it honors the structure. Put the format instruction in the Format part of your prompt so it isn't buried in context, and be specific about what you don't want too (e.g. "no preamble").
When should I use a Project instead of a one-off prompt?
Use a Project when you repeat work in the same context — a codebase, a client, a book, a brand. Add your background, files, style, and custom instructions as project knowledge, and Claude carries them into every chat inside the Project, so you stop re-explaining. Use a one-off prompt for anything you only do once.