Anthropic Prompt Library Walkthrough: Patterns You Can Reuse

The Anthropic prompt library is Anthropic's official set of copy-ready Claude recipes: system and user messages, tagged inputs, and clear output shapes for jobs like meeting notes, SQL help, citations, and support replies. You open an entry, study how it fences role from data from format, then rewrite those moves for your product instead of pasting someone else's sample text forever. This walkthrough shows how to browse the library, extract the patterns that still hold on Claude Fable 5, Claude Opus 5, Claude Sonnet 5, and Haiku 4.5, and adapt a starter into a prompt you own. Soft tip: PromptMake /text can draft the first adaptation pass at https://promptmake.net/text.

What the Anthropic prompt library is

Anthropic publishes curated prompts in its docs under Resources, often labeled Prompt Library. Each card names a job, shows a system prompt, a user prompt with placeholders, and often a sample turn. The point is teaching structure by example. You borrow the skeleton. You swap in your audience, your data, and your hard limits.

A second Anthropic surface sits beside it for Claude Code: short, outcome-first prompts tagged by discover, design, build, and review. That collection teaches agent habits (name the goal, attach artifacts, demand a check). Day-to-day Claude.ai and API users still mean the classic recipe cards when they search anthropic prompt library. Treat both as pattern banks, not as a finished team library.

Who this fits: PMs, support leads, analysts, and engineers who want Anthropic-shaped prompts without inventing XML tags from memory. Who should skip a pure browse: teams that already version production prompts in git with evals. Those teams need ownership and tests, which our internal prompt-library guide covers. Here the job is reading Anthropic's public recipes and turning one into yours.

PromptMake /text is the adaptation desk, not a mirror of Anthropic's site. You paste the job in plain language, target Claude, and get a tagged scaffold you edit. Guests get about three runs per day. Free accounts get about five. Keep confidential transcripts and customer IDs out of public tools when policy forbids third-party paste.

How Anthropic library entries are built

Open any strong library card and you see the same anatomy. A system block sets role, tone, and standing rules. A user block carries the variable material: transcript, table, ticket, or question. Tags or clear labels mark where the data starts and ends. An output rule names length, sections, or citation style. Examples appear when format drift would hurt.

That split exists because Claude weighs system guidance as durable behavior and treats the user message as the current ask. Library authors put stable persona and constraints in the system side. They put the unique payload in the user side. When people paste one giant blob with no boundary, Claude blurs rules and source text. Library cards avoid that blur on purpose.

You do not need to memorize every card title. You need to recognize five reusable moves: role with boundaries, tagged source material, explicit output shape, optional few-shot pairs, and a stop rule for missing facts. Once you see those moves, a marketing card and a coding card teach the same craft. The rest of this section unpacks those moves with concrete edits you can make in under two minutes.

Role, system rules, and user payload

Library system prompts sound calm and specific. They name who Claude writes for and what it refuses. Weak rewrites inflate ego: "world-class," "award-winning," "genius." Strong rewrites sound like this: "You write for ops managers at mid-market SaaS. You do not invent compliance certifications. You flag unknowns instead of guessing."

Keep the user message for the changing piece. Paste the meeting transcript, the CSV sample, or the ticket body inside a tag such as <transcript> or <ticket>. Put the ask after the data when the payload is long. Claude gives late instructions heavy weight, so bury the question and you get a vague essay.

With one short sentence of input, you can keep a single user message. Add a system block once the job repeats weekly or once the rules grow past three bullets. Library cards that look "heavy" earn that weight because the authors expected reuse.

Tags, output shape, and examples

XML-style tags in the Anthropic prompt library are labels, not a formal schema. <instructions>, <context>, <document>, and <output_format> tell Claude which paragraphs are rules versus source text. Rename tags to match your content: <customer_email>, <schema>, <brand_voice>. Two levels of nesting is enough for most work.

Output shape belongs in its own line or tag. "Three bullets under 20 words," "JSON with keys summary and next_step," or "Sections: Decisions, Owners, Open questions" beats another paragraph of personality. Add one or two short input-output examples when retries must match a format. Claude mirrors demonstrated patterns with less drift than style lectures alone.

Skip heavy markup for a one-line translation. Save long policy walls for production APIs you version in git. Day-to-day Claude work lives in short tagged blocks you can read in one scroll.

Step-by-step: adapt a library card into your prompt

Use the library as a pattern lab, then leave with a prompt that names your product. The loop below takes one card, one real task, and one edit pass. You finish with text you can paste into Claude.ai or the API and save in your notes.

Pick a job you will run this week: meeting notes, SQL explanation, support reply, or research brief with citations. Find a library card in the same family. Read the system and user blocks once for structure. Ignore cute sample names. Copy the skeleton into a scratch pad, then replace every placeholder with your facts. If the blank page still stalls you, dump the rough intent into PromptMake /text with Claude as the target and merge the generator scaffold with the library moves you liked.

Measure success by a usable first reply after one edit, not by collecting twenty unread cards. Three adapted templates beat a folder of untouched Anthropic samples.

Step 1: Choose the card by job, not by title charm

Match the deliverable. Need decisions and owners from a call? Open a meeting or scribe-style card. Need safe SQL from a schema? Open a SQL expert card. Need answers with sources? Open a citations-style card. Fancy titles sell curiosity. Job match saves time.

Skim the system block for rules you must keep: honesty about unknowns, tone, banned inventions. Skim the user block for which variables it expects. List those variables on paper: audience, source paste, length, format, forbidden claims. That list becomes your edit checklist.

If two cards look close, pick the shorter one. You can always add a constraint. Cutting a novel persona is harder than adding one line about row limits.

Step 2: Rewrite placeholders, then generate a Claude scaffold

Replace sample companies, fake tables, and toy transcripts with your material. Keep the tag names if they still fit. Change tag names when your content is clearer under a new label. Cut role fluff that adds no rule. Add the one constraint Anthropic could not know: SLA language, product names, banned claims, max length.

When the rewrite still feels thin, open https://promptmake.net/text, paste a three-sentence brief of the job, choose Claude, and generate once. Keep library-inspired tags and output rules. Drop generator lines that fight your constraints. Soft tip: the tool drafts structure; you own product truth.

Strip secrets before any public paste. Use redacted samples for structure tests. Move final runs with real customer text into approved Claude workspaces.

Step 3: Route the model and save the winner

Paste into Haiku 4.5 for fast drafts and format checks. Use Claude Sonnet 5 for everyday depth. Move the same tagged prompt to Claude Fable 5 or Claude Opus 5 when the lighter tier misses nuance twice on the same brief. Skip "think step by step" padding on reasoning-class models. Keep goals, evidence rules, and format crisp.

Save the final prompt next to the task name and the model you used. Label Haiku wins separate from Fable or Opus wins so you do not confuse lanes later. Next week, start from that file instead of browsing the Anthropic prompt library from zero.

If Claude misses a constraint, add one tagged sentence and resubmit. Avoid stuffing a new persona paragraph. Specific rules beat longer theater.

Patterns worth stealing from the Anthropic prompt library

Across cards, Anthropic repeats a small set of moves. Steal the moves. Leave the sample prose behind. The list below is the transfer kit you take into PromptMake /text or a blank notes doc.

Pattern A: Fence the source. Wrap transcripts, schemas, and emails in tags so Claude cannot treat them as instructions. Pattern B: Put the ask after long context. Pattern C: Name the reader and the banned moves in the system block. Pattern D: Demand a shape (bullets, table, JSON keys, memo sections). Pattern E: Add a missing-data rule: ask one clarifying question instead of inventing numbers.

Claude Code's library adds agent-side cousins of the same ideas: describe the outcome instead of micro-managing steps, attach the log or file, state a measurable check, and name the format for the answer. Those habits transfer to chat prompts when your job includes verify-and-retry language.

Worked rewrite: meeting notes card → your ops template

Library-style skeleton you might see: system role as a careful note-taker; user message with a <transcript> block; output sections for decisions, action items, and open questions.

Your adapted version: system rules that ban invented attendees and force "Unknown" when a owner is missing. User message with your real call paste inside <transcript>. Output format: Decisions (bullets), Owners (name + due date), Risks (max three), Open questions (max five). Add one line: "If the transcript never states a due date, write TBD."

Run on Claude Sonnet 5 first. If action items stay vague, move the same prompt to Claude Fable 5 and keep the tags. The upgrade is model depth, not a longer persona.

Worked rewrite: SQL help card → your analytics brief

Library-style skeleton: system role as a SQL specialist who explains queries; user message with schema plus a business question; output that includes the query and a plain-language explanation.

Your adapted version: paste the tables you allow Claude to touch inside <schema>. State dialect (Postgres, BigQuery, Snowflake). Add constraints: no DROP, no UPDATE, read-only SELECT, and "If a column is missing, ask instead of guessing names." Output format: SQL in a fenced block, then Assumptions, then Risks.

Haiku 4.5 often suffices for simple selects. Use Claude Opus 5 or Claude Fable 5 when joins span many tables and a wrong grain costs money. Keep the tagged schema identical across retries so model comparisons stay clean.

Mistakes that waste library time

Mistake 1: Copy-pasting sample companies and fake data into production chat. Readers notice. Models also learn the wrong nouns. Rewrite placeholders on the first use.

Mistake 2: Treating the Anthropic prompt library as a finished team library. Public cards lack your brand voice, SLAs, and owners. Save adapted winners where your team searches.

Mistake 3: Stacking CoT slogans on Claude Fable 5 or Claude Opus 5 after the card already set a clean goal. Extra scaffolding fights the model.

Mistake 4: Mixing Midjourney or FLUX parameter strings into a Claude library template. Wrong dialect. Retarget image work on PromptMake /image and keep text cards on Claude.

Mistake 5: Over-nesting tags four levels deep because a tutorial looked fancy. Flatten to two levels. Put the ask at the bottom.

Mistake 6: Browsing twenty cards and saving none. Adapt one job today. Collect later.

Mistake 7: Pasting secrets into a free generator your company forbids. Strip identifiers, or stay inside approved Claude workspaces.

Run this Anthropic library workflow today

Open the Anthropic prompt library and pick one card that matches a real task on your calendar. Copy the skeleton. Replace placeholders. Add audience, banned claims, and output shape. If you want a faster first draft of the tagged structure, run the rough idea once through https://promptmake.net/text with Claude selected, then merge.

Paste into Haiku 4.5, Claude Sonnet 5, or a Fable 5 / Opus 5 option based on difficulty. Save the final prompt with the model name. Repeat for two more job types this week. You now own a tiny personal library built from Anthropic patterns instead of a bookmark you never reopen.

When volume grows, register on PromptMake for a higher daily cap or move production prompts into versioned files your team controls. Until then, three careful adaptations beat thirty unread samples.

FAQ

What is the Anthropic prompt library?

It is Anthropic's public collection of Claude prompt recipes in the docs, plus related Claude Code starter prompts for coding workflows. Each classic entry shows system and user text you can copy and rewrite. The library teaches structure by example. You adapt the pattern for your data and keep the result where your team works.

Where do I find the Anthropic prompt library?

Start in Anthropic's Claude docs under Resources or Prompt Library, and check Claude Code docs for the coding-oriented prompt set. Search for the card that matches your job. Links move as docs reorganize, so use Anthropic's site search if a bookmark breaks. Prefer the official domain over random "prompt pack" mirrors.

How do I adapt an Anthropic prompt library card for my product?

Keep the role rules and output shape. Replace sample data with your transcript, schema, or ticket, and add constraints Anthropic could not know: product names, banned claims, length, dialect. Test on Haiku 4.5 or Claude Sonnet 5, then escalate to Claude Fable 5 or Claude Opus 5 if depth is thin. Save the edited prompt under your task name.

Should I use PromptMake /text with the Anthropic prompt library?

Use it when you want a fast Claude-tagged scaffold after you understand the library pattern. Paste a rough brief, choose Claude, generate once, then merge tags and output rules you took from the card. Soft link: https://promptmake.net/text. Skip the generator if your approved workspace already holds a strong template.

Is this the same as a Claude prompt generator guide?

A Claude prompt generator guide teaches a rough-idea-to-scaffold tool loop. This walkthrough teaches how to read Anthropic's official recipes and reuse their structure. You can learn patterns from the library, then draft variants with a generator. Keep the articles separate so search intent stays clear.

Which Claude model should I paste library-adapted prompts into?

Use Haiku 4.5 for speed and light formatting, Claude Sonnet 5 for daily depth, Claude Fable 5 for hard analysis, and Claude Opus 5 for heavy enterprise or coding workloads. Keep the same tagged prompt across lanes. Change the model when the lighter tier fails twice on the same brief.

Can I start adapting prompts without signing up for PromptMake?

PromptMake /text offers a small guest quota each day with no account. Create a free account if you need more daily runs. Check the live page for current numbers if quotas change. You can also adapt Anthropic cards by hand in a notes app with zero third-party tools.