Claude prompts that work: the anatomy and 15 examples by role
Claude prompts examples by role: 15 of them, the five slots a working prompt is built from, and how to graduate a good one into a Project or a Skill.
Ask for "ready-made Claude prompts" and you get a folder of text that does nothing on your task: it carries none of your context, none of your constraints, and nothing that tells you the output is good. Building your own takes two minutes. Below: the anatomy, a worked prompt, and 15 Claude prompt examples by role.
The anatomy of a prompt that works
Anthropic's prompt engineering guide opens with the part everyone scrolls past: you need success criteria and a way to test them before polishing wording. The rest falls into five slots.
Role. One line at the top focuses the answer, a technique in its own right in the docs.
Context. Who the client is, what the product does, how the last attempt ended. Anthropic recommends explaining why, not just what: not "never use ellipses", but "your response will be read aloud".
Constraints. Length, where facts may come from, what must not be invented.
Format. Say what to do, not what not to do: not "no markdown", but "smoothly flowing prose".
Done criteria. What separates a weak draft from a good one. Without it you edit answers instead of the prompt.
Two amplifiers: three to five varied examples, each in an <example> tag with the set in <examples>, plus XML tags like <instructions> and <context> for longer prompts.
Why magic phrases died
"You are a world-class expert, my job depends on it" belongs to the era when models struggled to hold an instruction. The docs now advise the opposite: recent models respond more strongly to the system prompt and can overtrigger on aggressive wording, so write plainly instead of "CRITICAL: you MUST". The same shift retired a workaround: from the 4.6 models on, a prefilled assistant response on the last turn returns a 400 error; earlier models still accept prefill.
One prompt in full, and what the second pass changed
*Role:* marketer at a grocery delivery service, writing a win-back email. *Context:* 412 people placed one order in 90 days and never came back; two out of three surveyed blamed a thin catalog, since grown from 4,000 to 11,000 items; the last discount campaign won back 3%. *Constraints:* under 900 characters, no discounts, no deadlines, no numbers beyond these. *Format:* subject under 50 characters, plain paragraphs, one button. *Done criteria:* it's clear what changed since that order, and a competitor couldn't send it under their own name.
The first answer was polite and empty: "we've improved, come back." The criterion caught it: any service could sign that line. On the second pass I added "the subject and first paragraph must carry a number from the context", and it opened with "4,000 items became 11,000". What carried it was the numbers and the discount ban, not the role, and a weak answer gets fixed in the prompt, not in the answer.
15 Claude prompt examples by role, and what carries each one
1. Marketer, offer. *Role:* marketer at a B2B service. *Context:* product, audience and three objections below. *Constraints:* 5 headlines under 60 characters, each hitting one objection. *Format:* one line under each naming the objection. *Criteria:* a competitor couldn't sign it. Without that: "make your business more efficient".
2. Marketer, competitors. "Four competitor pages below. Table: promise, price, proof. Where data is missing, say so." That permission keeps invented prices out.
3. Marketer, content plan. "Three segments below. 12 topics: topic, segment pain, format. Only what our cases back up." Evidence-tied topics drop the undeliverable ones.
4. Salesperson, objection. *Role:* account manager. *Context:* thread below, client went quiet after the price. *Constraints:* 3 messages under 400 characters, no pressure, no fake deadlines. *Format:* a line per option on the motive it uses. *Criteria:* you could send it to a friend without wincing. The limit and the banned tactics carry it.
5. Salesperson, call review. "Transcript below. Pull pain, budget, decision maker, next step, risks; mark what it lacks." Fixed fields beat "summarize the call".
6. Salesperson, proposal. "Structure: the client's problem in their words, solution, scope, timeline, price, next step. Numbers only from the brief." No invented numbers.
7. Manager, weekly digest. "Three reports below. One page: done, stuck, needs my decision. Every line has an owner and a date." Ownership makes it a to-do list.
8. Manager, decision. "Three options with cost of being wrong, what to verify first, who owns it. Recommend only after the third." The ban on an early verdict does the work.
9. Finance, margin. "Export below. Margin by line of business, formulas shown, plus every row you couldn't classify." The unclassified rows are the honest part.
10. Finance, scenarios. "Three quarterly scenarios: base, revenue minus 20%, plus 30%. For each, break-even and the month cash runs out, assumptions separately." Split-out assumptions expose the guesses.
11. HR, job post. "The five requirements below come from how the team works. Under 2000 characters, ending with month one duties." Month one forces specifics.
12. HR, resumes. *Role:* recruiter. *Context:* the role and 8 resumes below. *Constraints:* no personality judgments from the text, no guessed gaps. *Format:* per candidate, matches, gaps, two interview questions. *Criteria:* every conclusion pins to a line in the resume. Blocked guesses remove the "motivation" noise.
13. Legal, draft contract. "Find one-sided obligations and clauses with no deadline. Quote the clause, then comment. Do not rewrite." Quote-then-comment keeps it checkable.
14. Writer, article. "Topic, audience, what the reader should be able to do after. Outline of 6 sections, one claim and one example each. No intro about importance." The reader's outcome drives it.
15. Writer, edit pass. "Mark claims with no evidence and repeated ideas. Do not rewrite, list the fixes with quotes." Diagnosis split from treatment.
Graduating a prompt into a Project or a Skill
A prompt you paste a third time belongs in a tool.
Project. At claude.ai/projects: "+ New Project", then "Set project instructions" and "Save instructions", files via "+". One detail from the help center: context doesn't travel between chats unless it's in the knowledge base. Free caps at five projects. Assembling that base: projects and context.
Skill. A folder with a SKILL.md whose YAML frontmatter carries name (64 characters max) and description (200 max for a claude.ai upload, 1,024 in Claude Code and the API); Claude loads the skill off that description. The zip needs the skill folder at its root, not loose files. Anthropic's ready-made skills run on every plan including Free; your own zip needs Pro, Max, Team or Enterprise with code execution on: Settings > Capabilities, then Customize > Skills. On Team and Enterprise an owner enables it in Organization settings > Skills first, or the menu stays greyed out. The docs warn separately: a skill can hold executable scripts and touch your files, so install only your own or ones from a trusted source, and read a stranger's zip first.
Claude Code. In the terminal they're files: ~/.claude/skills/<name>/SKILL.md for personal, .claude/skills/<name>/SKILL.md for a project, invoked with /name. It needs a Pro, Max, Team or Enterprise plan or a Console account. Terminal Claude edits files on your machine and has the same network access as any other program, so read a skill before you run it. More on what Claude Code gives someone who doesn't code.
Common failures and honest limits
- No done criteria. The answer reads smooth with nothing to check it against.
- Data at the bottom. For large inputs, 20k tokens and up, the docs put documents above your query.
- One prompt for five jobs. The documented pattern: draft, review against criteria, revise.
- The skill never triggers. Check
descriptionand the zip layout: a vague description gives no signal on when to load it. - Expecting data from a prompt. The model can't see your CRM until you connect it. Personal skills in
~/.claude/skillslive on your machine: cloud sessions and Cowork load what's enabled on claude.ai. - Skipping the review. Numbers, quotes and anything going to a client or a contract get checked by hand: in legal and financial text a prompt saves time on the draft, never on the review. Interfaces move too, prefill being the proof.
The short version
Memorizing other people's prompts is pointless; your own takes two minutes: role, context, constraints, format, done criteria. Once one works, move it into a Skill or a Project.
At AGINE Academy you practice on your own tasks: missions run in the real Claude and every lesson ends with a working result. Four starter lessons are open without signup.
See also
Questions
Other people's packs are useful as a structure checklist: see which slots they fill and compare against yours. Past that they stall, because they carry none of your product, audience or criterion of good. Build your own from role, context, constraints, format, done criteria. Answers still vary run to run, but the quality becomes repeatable rather than accidental once you check every pass against the criterion.
Five slots. A one-line role, context (who the client is, what the product does, how the last attempt ended), constraints (length, where facts may come from, what must not be invented), format phrased as "do this" rather than "do not do that", and done criteria. Anthropic's docs also recommend explaining why, not only what, and wrapping each sample in an <example> tag with the whole set inside <examples>.
A one-line role helps; pressure doesn't. Recent models respond more strongly to the system prompt and can overtrigger on aggressive wording, so the docs suggest dropping "CRITICAL" and "you MUST" for plain language. The constraint and the criterion do the work. One old workaround is gone too: prefilling the last assistant turn returns a 400 error from the 4.6 models on, though earlier models still accept it.
A Project fits when shared material matters: instructions plus files in the knowledge base, where briefs, price lists and past copy live. One catch: context doesn't travel between chats in a project, it only comes from that knowledge base. A Skill fits when a repeatable procedure matters and you want to call it by name: a folder with SKILL.md, name and description in the YAML frontmatter. Uploading your own zip to claude.ai needs Pro, Max, Team or Enterprise with code execution on, and install third-party skills only from a source you trust, since they can carry executable scripts.
Numbers, quotes, links, and any fact heading to a client, a contract or a report. A prompt doesn't create data: if your internal figures aren't in the context, the model can't see them, and a plausible invention slides into the gap. The classic case: you ask for a conversion rate, forget to supply the order count, and the answer states a tidy "about 2.5%" that came from nowhere. In legal and financial text the saving happens on the draft, never on the review.