Claude Code Integration

brunnr skills work as custom slash commands in Claude Code. Install a skill, and it becomes a command your agent can invoke.

How it works

Claude Code supports custom slash commands through Markdown files in specific directories. brunnr skills are SKILL.md files -- Markdown documents with YAML frontmatter that serve as system prompts.

When you install a skill with brunnr install, it writes the SKILL.md to skills/<slug>/SKILL.md. To use it as a Claude Code slash command, you can either:

  1. Reference it directly -- Point Claude Code at the skill file.
  2. Copy to Claude Code's command directory -- Place it where Claude Code discovers custom commands.

Setup

Option 1: Project-level commands

Copy or symlink installed skills to your project's .claude/commands/ directory:

# Install the skill
brunnr install ax-rubric

# Create the commands directory
mkdir -p .claude/commands

# Symlink or copy
cp skills/ax-rubric/SKILL.md .claude/commands/ax-rubric.md

Now you can use it in Claude Code:

/ax-rubric "Query the knowledge base for structured learnings."

Option 2: Global commands

For skills you want available across all projects, use Claude Code's global command directory:

mkdir -p ~/.claude/commands
cp skills/ax-rubric/SKILL.md ~/.claude/commands/ax-rubric.md

Workflow

The typical workflow combines brunnr's security scanning with Claude Code's skill execution:

1. Browse and install

# List available skills
brunnr install --list

# Install with review
brunnr install ax-rubric

2. Scan before using

# Always scan skills before giving them to your agent
brunnr scan skills/ax-rubric/SKILL.md
=== brunnr scan ===

  CLEAN  SKILL

  Scanned 1 files: 1 clean, 0 flagged, 0 blocked

3. Deploy to Claude Code

cp skills/ax-rubric/SKILL.md .claude/commands/ax-rubric.md

4. Use in your agent session

/ax-rubric

Security considerations

Scan before you deploy

A SKILL.md file becomes a system prompt for your agent. If it contains malicious instructions (prompt overrides, credential theft patterns, data exfiltration commands), the agent will follow them. Always run brunnr scan before deploying a skill to Claude Code.

Defense in depth

  1. brunnr scan catches known attack patterns (7+1 threat classes)
  2. Review-before-install lets you read the skill content before writing it to disk
  3. brunnr eval validates that the skill produces expected output
  4. Manual review remains essential for novel attack patterns the scanner hasn't seen

Writing skills for Claude Code

When writing skills intended for Claude Code, keep these points in mind:

  • Use standard Markdown -- Claude Code renders the skill as a system prompt. Stick to headers, lists, tables, and code blocks.
  • Be explicit about phases -- Sequential instructions work better than unstructured prose.
  • Specify the output format -- Tell the agent exactly what to produce (tables, scores, rewrites, etc.).
  • Include trigger conditions -- "Use when reviewing tool descriptions" is better than "use when relevant."

See Writing Skills for the full guide.