AGENTS.md

Project Overview

• Name: Claude Code Prompts
• Type: Prompt library (Markdown-first) plus tested scripts bundled as Claude Code and cross-agent skills (.claude/skills/, .agents/skills/) and safety hooks
• Purpose: Production-ready system prompts for Claude Code, with cross-agent discovery through AGENTS.md, .agents/skills/, and llms.txt
• Core Methodology: APEI cycle — Analyze → Plan → Execute → Iterate
• License: MIT
• Language: English only

Routing skill

.claude/skills/find-prompt/SKILL.md is the canonical routing table. .agents/skills/find-prompt/SKILL.md is a thin cross-agent adapter. Keep the canonical table, indexes, llms.txt, and eval cases in sync whenever a prompt is added, removed, or renamed.

Repository Structure

├── AGENTS.md              # This file — project memory
├── README.md              # Main catalog, portfolio table, common combinations
├── REPOSITORY-MAP.md      # 30-second orientation, task -> file lookup
├── QUICK-START.md         # 30-second setup guide
├── USAGE.md               # Scenario-based prompt composition
├── CONTRIBUTING.md        # Contribution guidelines
├── CHANGELOG.md           # Version history (current: 2.2.0)
├── llms.txt               # Full LLM router index
├── .claude-plugin/
│   ├── plugin.json        # Claude Code plugin manifest
│   └── marketplace.json   # Installable marketplace catalog
├── hooks/
│   ├── hooks.json          # PreToolUse wiring
│   └── scripts/            # block-destructive-commands.sh, block-secret-writes.sh
├── .claude/skills/        # 6 Claude Code skills
│   ├── find-prompt/
│   ├── deterministic-checks/
│   ├── changelog-from-commits/
│   ├── doc-link-audit/
│   ├── skill-audit/
│   └── capability-audit/
├── .agents/skills/        # Cross-agent adapters / mirrored deterministic skills
├── evals/                  # find-prompt routing-accuracy regression tests (static + live tiers)
├── .github/workflows/
│   └── quality-gate.yml    # CI: lint, link audit, skill audit, deterministic-checks, plugin validate, routing eval
└── prompts/
    └── english/
        ├── INDEX.md           # Global task -> file router
        ├── agents/            # Active agent prompts + compatibility stubs
        │   ├── INDEX.md       #   catalog with token counts + task router
        │   └── archive/       # Archived prompts removed from active catalog
        ├── base/              # Foundation prompt (universal best practices)
        ├── project-types/     # Domain-specific prompts (11 files)
        ├── examples/          # Real-world usage examples
        └── workflows/         # Model selection, native features, Agent SDK,
            └── INDEX.md       #   APEI, selection, troubleshooting, setup, maintenance

Claude Code operation prompts (agents/)

The library covers the current Claude Code / Claude ecosystem. Keep these accurate against code.claude.com/docs and platform.claude.com/docs:

• agent-skills-prompt.md — SKILL.md authoring
• mcp-integration-prompt.md — MCP servers, scopes, auth, injection safety
• claude-code-plugins-prompt.md — plugin.json, marketplaces
• multi-agent-orchestration-prompt.md — native subagents + dynamic workflows
• hooks-automation-prompt.md — hook events and settings.json schema
• claude-code-workflow-prompt.md — CLAUDE.md, rules, settings, permissions
• claude-code-modes-prompt.md — adaptive thinking, effort, plan mode
• workflows/model-selection-guide.md — model lineup and effort
• workflows/claude-code-native-features-guide.md — the current feature surface
• workflows/agent-sdk-guide.md — the Claude Agent SDK

File Naming Conventions

Prompt Structure Standards

Every prompt follows this template:

# Prompt Title

> **Key Feature 1** | **Key Feature 2** | **Key Feature 3**

## Role
[Define what this prompt does]

## Protocol / Core Loop
[Main workflow — often uses a protocol acronym]

## Phases
[Phase details with templates and checklists]

## Remember
[Key takeaways — always the final section]

Key elements:

• Protocol acronym: Each prompt defines a memorable acronym for its core loop
• Phases: Detailed steps with Markdown templates and - [ ] checklists
• Tables: Used for options, comparisons, and structured data
• Code blocks: Used for command examples, templates, and diagrams
• Remember section: Closing callout with priority rules (always last)

Markdown Style Guide

• Use # for title, ## for major sections, ### for subsections
• Tables for structured data (catalogs, options, comparisons)
• > **bold text** for callout/subtitle lines under the title
• Fenced code blocks with language tags where applicable
• - [ ] checklists inside phases
• --- horizontal rules to separate major sections
• Keep prompts concise — include token count in catalog tables
• Use ⭐ emoji only for recommended/primary prompts in catalogs

Build & Development

No build system for the prompts themselves. The bundled skill/hook scripts under .claude/skills/*/scripts/ and hooks/scripts/ are standalone (bash/Python, no install step, no dependency beyond what's noted in each SKILL.md).

• Install: git clone the repo, or claude --plugin-dir <path> to try it as a plugin
• Validate links: python3 .claude/skills/doc-link-audit/scripts/check_links.py . — real link/anchor checker, not a grep spot-check
• Validate formatting: Open .md files in a Markdown previewer
• Lint: npx markdownlint-cli2 '**/*.md'
• Pre-commit scan: bash .claude/skills/deterministic-checks/scripts/scan.sh .
• Skill quality check: python3 .claude/skills/skill-audit/scripts/audit.py .claude/skills
• Plugin checks: claude plugin validate .claude-plugin/plugin.json and claude plugin validate .claude-plugin/marketplace.json

Bundled skills and scripts

.claude/skills/ ships 6 skills: find-prompt, 4 deterministic repository checks, and capability-audit. .agents/skills/ exposes the library to other agent hosts without maintaining a second routing table. hooks/ ships 2 PreToolUse safety scripts. Keep the catalogs, adapters, routing cases, and scripts synchronized; test scripts against real input before documenting them as working.

Common Tasks

Add a new agent prompt

1. Create prompts/english/agents/your-topic-prompt.md using the template above; open with a "Use this when" line and a "Skip to" anchor list
2. Update prompts/english/agents/INDEX.md — the task router and the catalog table (with a ~token estimate)
3. Update prompts/english/INDEX.md task router
4. Add a row to .claude/skills/find-prompt/SKILL.md and a routing eval case
5. Add entry to the portfolio table and, if it is a recurring pairing, the Common Combinations table in README.md
6. Add a line to llms.txt and REPOSITORY-MAP.md

Add a new project-type prompt

1. Create prompts/english/project-types/your-domain-prompt.md
2. Update prompts/english/INDEX.md
3. Add entry to the Foundation & Project Prompts table in README.md
4. Add relevant combinations to the Common Combinations table in README.md

Add a usage example

1. Create a new file in prompts/english/examples/
2. Show real-world prompt usage with context and expected output

Update documentation

1. Edit the relevant .md file
2. Ensure all relative links still resolve
3. Keep CHANGELOG.md updated for notable changes

Archive overlapping prompts

1. Classify prompt as keep / merge / archive
2. Move archived prompt to prompts/english/agents/archive/
3. Add archive rationale in prompts/english/agents/archive/INDEX.md
4. Update active catalogs (README.md, prompts/english/agents/INDEX.md, prompts/english/INDEX.md)

Quality Checklist

Before submitting changes:

• Content is accurate and follows best practices
• Markdown renders correctly (check headers, tables, code blocks)
• No spelling or grammar errors
• INDEX files updated if adding/removing prompts
• All relative links resolve to existing files
• Follows existing naming conventions (kebab-case-prompt.md)
• New prompts include Role, Protocol, Phases, and Remember sections
• README.md catalog tables updated if adding new prompts
• Recommendations are concrete (no vague advice)

Commit Messages

feat: add security audit prompt       # New features
fix: correct typo in API prompt        # Fixes
docs: improve README examples          # Documentation
update: enhance code review checklist  # Updates

Things to Avoid

• Don't add non-English content — this repo is English only
• Don't add code beyond the bundled skill/hook scripts (.claude/skills/*/scripts/, hooks/scripts/) — the prompt content itself stays pure Markdown; a script added there must ship inside a skill with a SKILL.md documenting it, be dependency-free or note its dependency explicitly, and be tested against real input before being documented as working
• Don't deviate from the APEI methodology in prompt design
• Don't remove the Remember section from prompts
• Don't use absolute URLs for internal links — use relative paths
• Don't cite Claude Code features, commands, or model IDs from memory — verify against code.claude.com/docs / platform.claude.com/docs
• Don't reference retired models (claude-3-*, claude-opus-4-1, claude-sonnet-4-0, gpt-4o) as current
• Don't add filler or AI-cliché phrasing; every sentence carries information