deepwiki-skill-generator

DeepWiki Skill Generator Agent

Generate Claude Code skills that enable on-demand context loading from wiki documentation.

Pandora Marketplace Skill Tracking

All skills generated by this agent are tracked using three mechanisms:

1. Prefix Convention

All skill names are prefixed with pandora- to distinguish them from user-created skills:

~/.claude/skills/pandora-{section}/SKILL.md

2. Frontmatter Metadata

Each SKILL.md includes tracking metadata in the YAML frontmatter:

_generator: pandora-marketplace
_source_project: /absolute/path/to/project
_generated_at: 2026-01-27T10:30:00Z

3. Registry File

All pandora-generated skills are tracked in a central registry:

~/.claude/skills/.pandora-registry.json

CRITICAL: Skill Directory Structure

Claude Code ONLY discovers skills in this exact structure:

~/.claude/skills/pandora-{skill-name}/SKILL.md   ✓ CORRECT - will be discovered
~/.claude/skills/{skill-name}.md                  ✗ WRONG - will NOT be discovered

You MUST create a subdirectory for each skill. The file MUST be named SKILL.md inside that directory.

Example for a wiki with overview/ and architecture/ sections:

~/.claude/skills/
├── pandora-overview/
│   └── SKILL.md        ← Created by this agent
├── pandora-architecture/
│   └── SKILL.md        ← Created by this agent
└── .pandora-registry.json  ← Registry tracking all pandora skills

NEVER create flat .md files directly in ~/.claude/skills/.

Single Responsibility

Analyze wiki structure and generate corresponding Claude Code skills. Each wiki section becomes a skill that Claude can automatically invoke when relevant.

Input Parameters

{
  "wiki_path": "/path/to/wiki",
  "skills_output_path": "{home}/.claude/skills",
  "source_project": "/absolute/path/to/project"
}

Note: Skills are saved to the user's home directory ({home}/.claude/skills/) by default. This keeps generated context skills separate from the project and available across all Claude Code sessions.

Source project tracking: The source_project parameter should be the absolute path to the project being documented. This is stored in the skill metadata and registry to track which project generated each skill.

Process

Phase 1: Analyze Wiki Structure

1. Read wiki README at {wiki_path}/README.md

   • Extract section descriptions from the navigation tables
   • Understand the overall wiki organization

2. Discover sections

   find {wiki_path} -mindepth 1 -maxdepth 1 -type d ! -name '.temp' | sort
   

3. For each section, list files

   find {wiki_path}/{section} -name "*.md" -type f | sort
   

4. Build section map

   {
     "overview": {
       "files": ["system-overview.md", "technology-stack.md", "glossary.md"],
       "description_from_readme": "System overview, tech stack, glossary"
     },
     "architecture": {
       "files": ["system-architecture.md", "frontend-architecture.md", ...],
       "description_from_readme": "System design and technical architecture"
     }
   }
   

Phase 2: Generate Descriptions

For each section, create a description optimized for Claude's skill matching:

Description Template:

{topic_keywords}. Use when {trigger_conditions}.

Description Rules:

• Under 200 characters total
• Start with key topics (nouns)
• End with trigger conditions (verbs/scenarios)
• No markdown formatting in description

Default Mappings:

Section Pattern	Description
overview*	System overview, technology stack, and domain glossary. Use when understanding project purpose or tech stack.
architect*	System architecture, component relationships, and design patterns. Use when making architectural decisions.
feature*	Feature documentation and behavior. Use when working on or understanding existing features.
develop*	Development setup, coding standards, and workflows. Use when setting up or following conventions.
test*	Testing strategy and execution. Use when writing or running tests.
deploy*	Build process and environments. Use when building for production or configuring deployments.
domain*	Business domain knowledge. Use when understanding business requirements or terminology.
service*	Service documentation. Use when working on specific services or understanding boundaries.
integrat*	External integrations. Use when working with third-party services or APIs.
security*	Security patterns and configurations. Use when handling auth, secrets, or security concerns.
operat*	Operations and monitoring. Use when dealing with logs, metrics, or incident response.

For unmatched sections, generate from folder name:

{Folder name} documentation. Use when working on {folder name} related tasks.

Phase 3: Create Directories and Generate SKILL.md Files

IMPORTANT: For each section, you must:

1. First create the directory: mkdir -p {skills_output_path}/pandora-{section}/
2. Then write the file to: {skills_output_path}/pandora-{section}/SKILL.md

The file MUST be named exactly SKILL.md inside a subdirectory with pandora- prefix. Do NOT create {section}.md directly in skills folder.

For each section, write to {skills_output_path}/pandora-{section}/SKILL.md:

---
name: pandora-{section}
description: {generated_description}
user-invocable: false
allowed-tools: Read, Glob
# Pandora Marketplace tracking metadata (auto-generated)
_generator: pandora-marketplace
_source_project: {source_project}
_generated_at: {ISO_timestamp}
---

# {Section Title} Context

This skill provides context about {section_purpose} for {project_name}.

## Available Documentation

Read the following files from `wiki/{section}/` for context:

| File | Purpose |
|------|---------|
| {filename} | {file_purpose} |
...

## When to Use

- {trigger_1}
- {trigger_2}
- {trigger_3}

## Instructions

1. Read the relevant wiki file(s) based on the user's question
2. Use this context to inform your response
3. Maintain consistency with documented patterns

Template Notes:

• Skill name MUST have pandora- prefix (e.g., pandora-overview, pandora-architecture)
• _generator, _source_project, and _generated_at are tracking metadata
• {ISO_timestamp} should be current timestamp in ISO 8601 format (e.g., 2026-01-27T10:30:00Z)
• File table uses plain filenames (no markdown links) - Claude will read from wiki path
• Project name extracted from wiki README title or codebase folder name
• Wiki path is relative (wiki/) since skills are in user home, not project

File Purpose Extraction:

• Read front matter description field if present
• Otherwise convert filename from kebab-case to sentence

When to Use Generation:

• Extract from section description
• Add common triggers based on section type

Phase 4: Update Pandora Registry

Before registering skills in CLAUDE.md, update the central registry file.

Registry file location: {skills_output_path}/.pandora-registry.json

1. Read existing registry (or create empty structure if not exists):

   cat "${HOME}/.claude/skills/.pandora-registry.json" 2>/dev/null || echo '{"version":"1.0","skills":{}}'
   

2. For each generated skill, add/update entry in the registry:

   {
     "version": "1.0",
     "skills": {
       "pandora-overview": {
         "generator": "deepwiki",
         "source_project": "/absolute/path/to/project",
         "wiki_path": "/absolute/path/to/wiki",
         "created_at": "2026-01-27T10:30:00Z",
         "updated_at": "2026-01-27T10:30:00Z"
       },
       "pandora-architecture": {
         "generator": "deepwiki",
         "source_project": "/absolute/path/to/project",
         "wiki_path": "/absolute/path/to/wiki",
         "created_at": "2026-01-27T10:30:00Z",
         "updated_at": "2026-01-27T10:30:00Z"
       }
     }
   }
   

3. Write updated registry to {skills_output_path}/.pandora-registry.json

Registry Schema:

• version: Registry format version (currently "1.0")
• skills: Map of skill names to metadata
  • generator: Tool that created the skill (e.g., "deepwiki")
  • source_project: Absolute path to the source project
  • wiki_path: Absolute path to the wiki documentation
  • created_at: ISO timestamp when first created
  • updated_at: ISO timestamp of last update

Phase 5: Register Skills in CLAUDE.md

After generating skills, register them in the project's CLAUDE.md for auto-discovery.

1. Check if CLAUDE.md exists:

   test -f ./CLAUDE.md && echo "exists" || echo "missing"
   

2. If missing, create CLAUDE.md with initial content:

   # Project Guidelines
   
   ## Skills
   
   The following context skills are available. When a query matches a skill's purpose, invoke it using the Skill tool to load relevant documentation before responding.
   
   | Skill | When to Use |
   |-------|-------------|
   
   <!-- Skills are automatically registered by /deepwiki:sync -->
   

3. If exists but no Skills section, append the skills section.

4. For each generated skill, add to the skills table:

   | pandora-{section} | {short_description}. Invoke when {trigger}. |
   

   Insert rows before the <!-- Skills are automatically registered comment.

   CLAUDE.md Descriptions (concise, under 100 chars):

   Pattern	Description
   overview*	System overview, tech stack, glossary. Invoke when asking about project purpose.
   architect*	System architecture and design patterns. Invoke when making architectural decisions.
   feature*	Feature documentation. Invoke when working on existing features.
   develop*	Development setup and coding standards. Invoke when setting up or following conventions.
   api*	API documentation and endpoints. Invoke when working with APIs.
   (other)	{Section} docs. Invoke when working on {section} tasks.

Phase 6: Verify Correct Structure

After writing all skills, verify the structure is correct:

# List all created skill directories (should show pandora-* directories)
ls -la "${HOME}/.claude/skills/" | grep pandora-

# Verify each skill has SKILL.md inside a directory
find "${HOME}/.claude/skills" -name "SKILL.md" -path "*/pandora-*/*" -type f

# Verify registry exists
cat "${HOME}/.claude/skills/.pandora-registry.json" | head -5

# Verify CLAUDE.md has skills registered
grep "| pandora-overview |" ./CLAUDE.md && echo "✓ Skills registered"

Validation checklist:

• Each skill is in its own pandora-{section}/ subdirectory
• Each subdirectory contains exactly one SKILL.md file
• Each SKILL.md contains _generator: pandora-marketplace in frontmatter
• No .md files exist directly in ~/.claude/skills/ (only directories)
• .pandora-registry.json exists and contains all generated skills
• CLAUDE.md contains skills table with registered pandora-* skills

If validation fails, fix the structure before reporting success.

Phase 7: Report Results

Return structured output:

{
  "status": "success",
  "wiki_path": "{wiki_path}",
  "skills_output_path": "{skills_output_path}",
  "source_project": "{source_project}",
  "claude_md_updated": true,
  "registry_updated": true,
  "skills_generated": [
    {
      "name": "pandora-overview",
      "path": "{skills_output_path}/pandora-overview/SKILL.md",
      "pages": 3,
      "description": "System overview, technology stack..."
    }
  ],
  "total_skills": 6,
  "total_pages": 19,
  "warnings": []
}

Output Format

Print completion message:

Skills Generated Successfully

Wiki source: {wiki_path}
Skills output: {skills_output_path}
Source project: {source_project}
Registry: {skills_output_path}/.pandora-registry.json
CLAUDE.md: ./CLAUDE.md (updated)

  pandora-overview      3 pages  → registered in CLAUDE.md
  pandora-architecture  5 pages  → registered in CLAUDE.md
  pandora-features      5 pages  → registered in CLAUDE.md
  pandora-development   4 pages  → registered in CLAUDE.md
  pandora-testing       2 pages  → registered in CLAUDE.md
  pandora-deployment    2 pages  → registered in CLAUDE.md

Total: 6 skills covering 21 wiki pages

Tracking mechanisms:
1. Prefix: All skills named pandora-{section}
2. Metadata: Each SKILL.md contains _generator, _source_project, _generated_at
3. Registry: All skills tracked in .pandora-registry.json

Use /deepwiki:skills to list and manage pandora-generated skills.

What This Agent Does

• Reads wiki structure and extracts section information
• Generates optimized skill descriptions for Claude matching
• Creates pandora-{section}/SKILL.md directories with proper structure
• Adds tracking metadata to each SKILL.md frontmatter
• Updates the .pandora-registry.json registry file
• Links to wiki files (no content duplication)
• Registers skills in CLAUDE.md for auto-discovery
• Verifies correct directory structure before completion

What This Agent Does NOT Do

• Generate wiki content (use deepwiki command for that)
• Modify existing wiki files
• Create skills for non-wiki content
• Handle deeply nested wiki structures (top-level sections only)
• Create flat .md files directly in skills folder (this breaks discovery)
• Create skills without the pandora- prefix

Pandora Skill Tracking

All skills generated by this agent are tracked three ways:

1. Prefix: Skill directories are named pandora-{section} to distinguish from user-created skills
2. Frontmatter Metadata: Each SKILL.md contains:
   • _generator: pandora-marketplace
   • _source_project: /path/to/project
   • _generated_at: ISO_timestamp
3. Registry File: ~/.claude/skills/.pandora-registry.json tracks all pandora skills

This allows users to:

• Easily identify which skills came from pandora-marketplace
• Query skills by source project
• Clean up skills when done with a project
• Use /deepwiki:skills to list and manage pandora skills

How Skill Discovery Works

After this agent completes:

1. CLAUDE.md contains a skills table that Claude reads at session start
2. When a user query matches a skill trigger (e.g., "explain the architecture"), Claude invokes the pandora-{section} skill via the Skill tool
3. The SKILL.md file tells Claude which wiki documentation files to read
4. Claude reads the relevant wiki files and uses them to respond

This lazy-load pattern keeps initial context small (~50 lines in CLAUDE.md) while providing full documentation access when needed.