@a16njs/plugin-claude

Claude Code plugin for a16n. Discovers and emits Claude configuration.

Installation

This plugin is bundled with the a16n CLI. For programmatic use:

npm install @a16njs/plugin-claude

Supported Types

This plugin supports five customization types:

Note: Claude has no dedicated command concept. AgentCommands from Cursor are emitted as skills with a description enabling /command-name invocation. The Claude plugin never discovers AgentCommands—conversion is one-way (Cursor → Claude only).

Supported Files

Discovery

• CLAUDE.md - Root Claude configuration (GlobalPrompt)
• */CLAUDE.md - Nested Claude configuration files (GlobalPrompt)
• .claude/skills/*/SKILL.md - Skills with description frontmatter (AgentSkill)
• .claude/settings.json - Permissions deny rules (AgentIgnore)

Note: Skills with hooks: in their frontmatter are skipped (not convertible to Cursor). Note: Only Read() permission denials are discovered (other types like Bash() or Edit() are ignored).

Emission

• GlobalPrompt → CLAUDE.md (merged with section headers)
• FileRule → .a16n/rules/<name>.txt + .claude/settings.local.json with hooks
• AgentSkill → .claude/skills/<name>/SKILL.md with description frontmatter
• AgentIgnore → .claude/settings.json with permissions.deny Read rules
• AgentCommand → .claude/skills/<commandName>/SKILL.md with Invoke with /command description

File Formats

CLAUDE.md (GlobalPrompt)

# Project Guidelines

Your instructions for Claude here.

SKILL.md (AgentSkill)

---
description: Testing best practices
---

Write unit tests first.
Aim for 80% code coverage.

settings.local.json (FileRule via hooks)

FileRules are converted using @a16njs/glob-hook for runtime glob matching:

{
  "hooks": {
    "PreToolUse": [{
      "matcher": "Read|Write|Edit",
      "hooks": [{
        "type": "command",
        "command": "npx @a16njs/glob-hook --globs \"**/*.tsx\" --context-file \".a16n/rules/react.txt\""
      }]
    }]
  }
}

Note: FileRule conversion emits an "Approximated" warning because hook-based matching may differ slightly from Cursor's native glob matching.

settings.json (AgentIgnore via permissions.deny)

AgentIgnore patterns are converted to permissions.deny Read rules:

{
  "permissions": {
    "deny": [
      "Read(./dist/**)",
      "Read(./.env)",
      "Read(./**/*.log)",
      "Read(./secrets/**)"
    ]
  }
}

Pattern conversion rules:

• dist/ → Read(./dist/**)
• .env → Read(./.env)
• *.log → Read(./**/*.log)
• **/*.tmp → Read(./**/*.tmp)

Note: AgentIgnore conversion emits an "Approximated" warning because Claude's permission system may behave slightly differently than .cursorignore.

SKILL.md from AgentCommand

When Cursor commands are converted to Claude, they become skills:

---
name: "review"
description: "Invoke with /review"
---

Review this code for:
- Security vulnerabilities
- Performance issues
- Code style violations

The description: "Invoke with /review" enables slash command invocation in Claude, mirroring Cursor's /review behavior.

Usage

import claudePlugin from '@a16njs/plugin-claude';
import { A16nEngine } from '@a16njs/engine';

const engine = new A16nEngine([claudePlugin]);

// Discover Claude configuration
const result = await claudePlugin.discover('./my-project');
console.log(`Found ${result.items.length} items`);

// Emit to Claude format
await claudePlugin.emit(result.items, './my-project');