6 effective Claude Code prompt patterns from Anthropic's Prompt library
TL;DR
Good Claude Code prompts state the outcome rather than steps, give a way to self-check, point at a reference, set a measurable target, include the original artefact and say how the answer should look. Save prompts that work repeatedly as skills.
With the same Claude Code, one person gets a good result on the first turn while another has to correct it five times. The difference is usually how the prompt is written. Anthropic has a Prompt library for Claude Code: copy-paste prompts tagged by task and role, each with a "why this works" note. The most valuable part isn't any single prompt but the six patterns behind them.
Six patterns that make prompts work
| Pattern | Example |
|---|---|
| Describe the outcome, not the steps | "add rate limiting to the public API and make sure existing tests still pass" |
| Give it a way to check its work | "write the migration, run it against the dev database, and confirm the schema matches" |
| Point at a reference | "add a settings page that follows the same layout as the profile page" |
| State the measurable target | "get the bundle size under 200KB and show me what you removed" |
| Give it the artefact | "why is the build failing? @build.log" (paste logs, screenshots, errors instead of describing them) |
| Say how you want the answer | "explain the payment retry logic as an HTML page with a diagram, then open it in my browser" |
Why each pattern works
- Outcome over steps: Claude finds the files itself instead of following the path you guessed.
- Self-checking: with a command, test or comparison in the prompt, Claude iterates until it passes rather than stopping after one attempt. To check against the running app, there's
/verify. - A reference: new code stays consistent with existing code, with fewer style fixes.
- Measurable: "done" becomes unambiguous rather than a debate.
- The original artefact: Claude reads the source rather than your description of it, so less is misunderstood.
- Answer format: the answer fits how you'll use it; for a lasting default, set an output style.
One prompt combining all six
The /checkout page is slow (@lighthouse.json). Get LCP under 2.5 seconds,
following the image optimisation already used on /product. Rerun Lighthouse to
confirm, and don't change the UI. Finally, summarise the changes as a
"file / change / impact" table.From a good prompt to a repeatable workflow
- A prompt that works repeatedly → save it as a skill so the team can run it as
/command. - Conventions Claude learns along the way → record them in CLAUDE.md so later sessions don't relearn them.
- Large or risky changes → use plan mode to see the file list before any edits.
Caveats
- Library prompts are starting points, not scripts: adapt them to your project.
- Don't dictate steps unnecessarily: detailed procedures make it harder for Claude to find a better way.
- Artefacts can contain secrets: scrub tokens and passwords from logs before pasting.
How to apply it
- Dev: Before sending a prompt, ask: is there a way to check, and did I point at an existing example?
- Tech Lead: Browse the Prompt library, pick 5–10 prompts that fit your team and turn them into internal skills.
- PM: Write requests as "outcome + measurable criteria + report format"; that's also a good prompt for Claude.
Sources & references
- Prompt library — Claude Code Docs