Promptshelf Pattern Library · 40 patterns · 5 categories
Prompt Pattern Library · 40 reusable instruction shapes

Forty prompt patterns with side-by-side input and output.

Every pattern card pairs the prompt template you paste with a concrete illustrative response — so you can calibrate what the shape produces before you spend a turn on it. Five categories (research, code, debug, refactor, write). Filter by category, star your favorites, share your shortlist via URL.

40 / 40 visible 5 / 5 categories on 0 starred 40 patterns · 5 categories · zero accounts
Bay 01

Filter, star, and share.

Categories
Search
Showing all 40 patterns.
No patterns match these filters. Toggle a category on, clear the search, or disable "starred only."
Bay 02

How to use a pattern.

Each card has two panes. The Input pane is the prompt template you paste (substitute any placeholder in {braces} for your own content). The Output pane shows the shape of response the pattern aims to produce — a calibration reference, not a transcript. Click the star to add a pattern to your shortlist; the URL updates live so you can bookmark or share your curated set. Click Copy input to grab the template alone.

If you prefer worked before/after examples over abstract patterns, see our Before/After Vibe-Coded Gallery — 20 side-by-side pairs that use many of these same instruction shapes. For ready-to-paste Claude Code slash commands, see the Slash Command Snippet Library.

Bay 03

Compound these with our other tools.

Gallery
Before/After Vibe-Coded Gallery
20 worked pairs — vague prompt versus an engineered rewrite applying these patterns. Case studies for the grammar.
Quiz
AI Tool Quiz
Eight questions that surface the AI-coding stack (Claude Code, Cursor, Copilot, or a mix) that fits your workflow.
Guide
CLAUDE.md Guide
How to encode these patterns as persistent project instructions rather than pasting every session.
001 · Free ZIP
Claude Code Starter Kit
Skill templates, CLAUDE.md boilerplate, hooks cookbook, and settings.json profiles that bake several of these patterns into the session.
Bay 04

FAQ.

What counts as a prompt pattern?
A prompt pattern is a reusable instruction shape that biases the model toward a specific kind of output — chain-of-thought, schema-first, counter-example sweep. Each card here names the shape, shows a minimal input that applies it, and shows the expected output shape so you can calibrate before pasting.
Why five categories and not more?
Research, Code, Debug, Refactor, and Write cover the phases most knowledge-work sessions actually pass through. Every "new" category proposed during authoring collapsed into one of these five — "testing" is Code + Debug, "planning" is Research + Write, and so on. Five is also the natural bit-budget for a 1-hex filter mask (0x00–0x1f).
How does the share URL work?
Each category maps to one bit (research=0, code=1, debug=2, refactor=3, write=4). The five bits pack into one or two lowercase hex characters — ?f=1f means all five on, ?f=02 means only code. Starred cards are appended as ?s=<slug>,<slug>,<slug>. The URL updates live as you filter or star; copy it, share it, bookmark it.
Are the example outputs real LLM responses?
No. Every output is a hand-written illustrative shape — what a well-calibrated response to that prompt would look like. Real LLM responses vary by model, temperature, and context; committing to one captured response would rot as models update. The outputs here are calibration references, not transcripts.
How are these different from the Vibe-Coded Gallery?
The Gallery compares a vague prompt to an engineered rewrite of the same prompt — before/after pairs. The Pattern Library catalogs the reusable instruction shapes that the engineered side uses. Think of the Gallery as case studies and the Library as the grammar.
Is the page sending any data?
No. Filters, starring, and share-URL encoding all run in the browser. No analytics, no telemetry, no third-party scripts beyond Google Fonts (CSS-only). Your starred slugs live in the URL and optionally localStorage. Nothing leaves your device.
Can I adapt these for Cursor or Copilot?
Yes. The patterns are tool-agnostic — they describe instruction shapes, not Claude-specific APIs. For Cursor, paste into .cursor/rules/<slug>.md or into an Agent-mode system prompt. For Copilot, save as a .prompt.md file and load via workspace settings. The few Claude-Code-specific references (TodoWrite, Bash tool) have obvious Cursor/Copilot equivalents.
How do I tip the maker?
BTC: bc1qs04leape97ner4wqa98n94l9n0gv9aa84eg4ux (copy button in the tip jar below). No accounts, no signup, no middleman. If a pattern here saved you an hour, a few thousand sats is a fair thanks — but everything on Promptshelf stays free whether or not you tip.

Saved you an hour? Tip the maker in BTC — no account, no signup, just paste.

BTC bc1qs04leape97ner4wqa98n94l9n0gv9aa84eg4ux