Claude Prompts: Ready-Made Templates
In short
A strong Claude prompt isn't a magic phrase — it's a four-part structure: role (who is answering), context (what's given), task (what to do) and success criteria (how to tell it's done). Claude reads explicit markup especially well: wrap prompt blocks in XML tags like <context> and <task> — that's Anthropic's official recommendation. The main difference from GPT: Claude follows instructions literally, so ALL CAPS and "THIS IS MANDATORY!!!" don't strengthen a prompt, they break it. Below are 10 ready-made templates for code, review, writing, analysis and planning, plus a separate section on Claude Code. Another 79 prompt cards are free in the arsenal section.
What makes a strong Claude prompt?
A prompt is a spec for the model: the sharper the spec, the fewer revisions. The working frame has four required parts and one optional one:
- Role — who the model should be: "strict reviewer", "business copy editor". The role switches on the right layer of knowledge and the right tone.
- Context — everything that affects the result: product, audience, stack, constraints, source data.
- Task — one specific action with a specific output. Not "help me with this text", but "rewrite this paragraph at half the length".
- Criteria — what "done" looks like: format, length, tone, what's off limits. Their absence is the single most common cause of bad answers.
- Examples (optional). Few-shot means showing 2-3 "input → output" samples so the model copies the pattern. For repetitive tasks it beats any description.
Template 1. Universal frame
You are a [role: senior developer / editor / analyst].
<context>
[Product, audience, constraints, source data.]
</context>
<task>
[One specific task with a specific output.]
</task>
<criteria>
[Format, length, tone. What is off limits. How to tell it's done.]
</criteria>
For short requests XML tags are optional, but in prompts longer than a couple of paragraphs they visibly cut confusion: the model sees exactly where context ends and the task begins.
How do Claude prompts differ from GPT prompts?
About eighty percent of the techniques overlap: the "role → context → task → criteria" chain is universal. The differences are in the details, but they change the output:
| Aspect | Claude | GPT (ChatGPT) |
|---|---|---|
| Block markup | XML tags <context>, <example> — Anthropic's official recommendation |
Markdown headings and separators |
| Instruction style | Literal compliance: calm wording beats caps | More tolerant of intensifiers like CRITICAL and MUST |
| Where to set the role | First line of the prompt or in the project's system prompt | Custom Instructions / GPTs settings |
| Long documents | Document at the start, question at the end (Anthropic's recommendation) | No hard recommendation |
| Few-shot examples | 2-3 examples inside <example> tags |
Examples in free form |
A note on literalism: current Claude models (as of July 2026 — Sonnet 5 and Opus 4.8) follow instructions more precisely than previous generations. Anthropic says so plainly in its migration guides: prompts stuffed with "CRITICAL: YOU MUST" lead to overwork — the model does exactly what's written even when it shouldn't. Write the way you'd brief a competent colleague, without shouting.
Code templates: write it, review it, fix it
Template 2. New function or feature
You are a senior developer. Stack: [React + TypeScript / Python + FastAPI].
<task>
Write a [function/component/endpoint]: [what it does, what goes in, what comes out].
</task>
<criteria>
- Working code, no stubs and no "the logic would go here"
- Handle the edge cases: [empty input, network error, duplicates]
- At the end, 3 lines: how to check that it works
</criteria>
Template 3. Code review against a checklist
You are a strict reviewer. Check the code in this order:
security → logic errors → performance → readability.
For every finding: the line, what's wrong, how to fix it,
severity (high / medium / low). Don't rewrite the whole file.
If there are no problems, say so — don't invent findings.
<code>
[code]
</code>
An extended version with the full checklist lives in the "Code review against a checklist" card.
Template 4. Hunting a bug
The code is failing: [error text]. I expect: [what should happen].
I get: [what actually happens].
Give me 3-5 hypotheses for the cause, most likely first.
For each one, a way to check it in under a minute.
Don't propose a fix until a hypothesis is confirmed.
<code>
[code + stack trace]
</code>
Why "hypotheses, not a fix": a model told to "just fix it" patches the first suspicious line it sees. The technique is broken down in the "Find the bug" card.
Templates for writing and analysis
Template 5. Copy without AI clichés
You are an editor who hates bureaucratic filler and AI clichés.
<task>Write a [post/article/email] about [topic] for [audience].</task>
<criteria>
- Banned: "in today's world", "let's dive in", "it's no secret that"
- Short sentences. Numbers and examples instead of opinions
- One idea per paragraph. Length: [N] words
</criteria>
If the text is already written and "smells of a neural net", take 6 prompts that strip out AI style.
Template 6. Summarising a long document
Produce a structured summary of the document.
Format:
1. The gist — one paragraph
2. 5-7 key points, each with a supporting quote
3. All figures and dates as a separate list
4. What is still unclear
If something isn't in the document, write "not in the document". Don't fill in the blanks.
<document>
[text]
</document>
Template 7. Data analysis
You are an analyst. Data: [paste a CSV or a table].
1. Find 3-5 patterns that are genuinely visible in the data
2. For each one, the supporting numbers and a possible explanation
3. Flag the anomalies separately
4. Tell me which data is missing for reliable conclusions
Don't invent values. If the data is too thin, just say so.
Lines like "don't fill in the blanks" and "just say so" aren't paranoia: without them the model papers over gaps with plausible inventions. That's your main insurance against hallucinations in production prompts.
Planning templates
Template 8. Breaking down a task
Help me plan: [goal]. Deadline: [date]. Resources: [who and what I have].
1. Break it into stages. For each, the outcome in "done when…" form
2. A time estimate and the main risk for every stage
3. What to cut first if the deadline slips
4. Ask me the 3 questions that would sharpen the plan the most
Template 9. A week's content plan
You are a content strategist. Product: [what]. Audience: [who]. Platform: [where].
Build a plan for 7 posts. For each: format (post/carousel/video),
the first-line hook, 3 talking points, a call to action.
No repeated topics. Answer as a table.
The "ask me questions" line belongs in any planning prompt: Claude is good at exposing holes in how a task is framed, but by default it politely plans around them.
Prompts for Claude Code: what changes?
Claude Code is an agent in the terminal: it doesn't reply with text, it reads files, writes code and runs commands. The prompt is closer to a work order, and the emphasis shifts:
- Context lives in a file, not in the prompt. CLAUDE.md is the project's memory file: stack, build commands, prohibitions. Written once, applied in every session — which makes prompts shorter.
- Plan first, code second. Ask for a change plan and approve it: fixing a plan is cheaper than redoing an implementation.
- The done criterion must be checkable. "Done = tests pass" works better than "do it carefully".
Template 10. A task for Claude Code
Read src/auth/. Draft a plan: how to add password recovery.
Plan only for now, no code. Once I say "ok", implement it step by step.
Done = npm test passes and the new flow is documented in README.
The basics of the tool are in the Claude Code guide for beginners; the advanced level is subagents, where a task is split across parallel workers. We packaged these rules — plan before code, checkable criteria, project memory — into the Quest engine for Claude Code and Cursor: the scaffolding is already configured, so you don't have to re-prompt every session from scratch.
Where to find more ready-made prompts
The qvib arsenal holds 79 prompt cards, free: each with a template, a sample result and an honest breakdown of when the technique works and when it doesn't. There are technique cards (few-shot, chain-of-thought, persona) and curated sets — for example, 25 Claude prompts for business. Copy them, drop in your own specifics and compare the before and after.
FAQ
Do prompts written for ChatGPT work in Claude?
Mostly yes: the "role → context → task → criteria" structure is universal. When porting, strip the intensifiers — caps, "CRITICAL", triple exclamation marks: Claude executes them literally and the answer skews. Wrap long context blocks in XML tags.
Do I have to write prompts in English?
No. Claude handles other languages confidently — it understands and writes without quality loss. The exception is technical terms in coding tasks: library, function and pattern names are better left untranslated.
Is a longer prompt always a better prompt?
No. What works is density, not length: everything that affects the result goes in, everything that doesn't gets cut. Extra detail dilutes the task. A short prompt with sharp criteria beats a long one without them.
What if Claude answers the wrong way?
Check the criteria first: nine times out of ten the model missed the format because nobody specified it. Then show an example of the answer you want — few-shot fixes most "not quite that" cases. And ask Claude to improve your own prompt: "rewrite this prompt so the answer is [X]" — it's excellent at that meta-task.