skill-distillery

Skill Distillery

Transform external repositories into Claude Code plugins.

External Repo → Research → Recon → Patterns → Codify → Author → Package → Audit → Plugin

Steps

1. Clarify target repo and plugin goals
2. Load the research skill for external discovery (docs, APIs, community patterns)
3. Load the codebase-analysis skill for internal analysis of target repo
4. Load the find-patterns skill to extract repeatable patterns worth automating
5. Load the codify skill to map patterns to component types
6. Author components using skillcraft or claude-craft
7. Load the claude-plugins skill to package into distributable plugin
8. Validate using the claude-plugins audit checklists

<when_to_use>

• Turning a CLI tool into a Claude Code plugin
• Creating skills that wrap an external library or API
• Building plugin companions for MCP servers
• Extracting automation opportunities from a third-party repo
• Packaging workflow patterns around external tools

NOT for: plugins from scratch (use claude-plugins), single-skill creation (use skillcraft), auditing existing plugins (use claude-plugins)

</when_to_use>

Artifact Structure

Track progress with artifacts in artifacts/skill-distillery/:

artifacts/skill-distillery/
├── discovery.md      # Research output (docs, APIs, community patterns)
├── recon.md          # Codebase analysis (structure, conventions, key files)
├── patterns.md       # Extracted patterns with automation value
├── mapping.md        # Pattern → component mapping decisions
├── components/       # Authored skills, agents, hooks, commands
│   ├── skill-1/
│   ├── skill-2/
│   └── ...
└── audit.md          # Plugin validation results

Quick Mode

For simple repos (single-purpose CLI, small API wrapper):

1. Skip stages 3-4 — go direct from recon to authoring
2. Create 1-2 skills covering primary use cases
3. Package immediately

Trigger: User says "quick", repo has < 5 main commands, or clear single purpose.

Stages

Load the maintain-tasks skill for stage tracking. Stages advance only.

Stage	Skill	activeForm
1. Discovery	research	"Researching external docs"
2. Recon	codebase-analysis	"Analyzing target repo"
3. Patterns	find-patterns	"Extracting patterns"
4. Mapping	codify	"Mapping to components"
5. Authoring	skillcraft	"Creating components"
6. Packaging	claude-plugins	"Packaging plugin"
7. Audit	claude-plugins	"Validating plugin"

Stage 1: Discovery

Load research skill. Gather external docs, community patterns, pain points.

See stage-1-discovery.md for details.

Stage 2: Recon

Load codebase-analysis skill. Analyze structure, API surface, conventions.

See stage-2-recon.md for details.

Stage 3: Patterns

Load find-patterns skill. Extract workflows, command sequences, decision points.

See stage-3-patterns.md for details.

Stage 4: Mapping

Load codify skill. Map patterns to component types (skill, command, hook, agent).

See stage-4-mapping.md for details.

Stage 5: Authoring

Load appropriate skill per component type. Create in artifacts/skill-distillery/components/.

See stage-5-authoring.md for details.

Stage 6: Packaging

Load claude-plugins skill. Create plugin structure with manifest and README.

Ask: "Do you have an existing marketplace to add this plugin to?" If yes, prepare the marketplace entry.

See stage-6-packaging.md for details.

Stage 7: Audit

Validate using the claude-plugins audit checklists. See its references/audit.md.

See stage-7-audit.md for details.

<decision_points>

Key decisions during engineering process:

Which patterns to automate?

• High frequency + medium complexity = best ROI
• Low frequency + high complexity = consider if audience is technical
• One-off patterns = skip

Skills vs Commands?

• Multi-step, needs guidance → Skill
• Quick action, obvious usage → Command
• User entry point to skill → Both (command loads skill)

Include agents?

• Only for complex repos with orchestration needs
• Most plugins don't need custom agents
• Consider if existing agents (analyst, engineer) suffice

Quick mode vs full pipeline?

• Single-purpose tool → Quick mode
• Complex tool with many features → Full pipeline
• Unclear → Start with recon, decide after

</decision_points>

ALWAYS:

• Start with discovery before touching code
• Track artifacts at each stage
• Validate patterns have 3+ use cases before creating components
• Use existing skills for authoring (don't reinvent)
• Run audit before declaring complete

NEVER:

• Skip recon stage (even for familiar repos)
• Create agents without clear orchestration need
• Package without audit validation
• Over-engineer (simpler plugin > feature-complete plugin)

Stage guides:

• overview.md — quick reference
• stage-1-discovery.md — external research
• stage-2-recon.md — codebase analysis
• stage-3-patterns.md — pattern extraction
• stage-4-mapping.md — component selection
• stage-5-authoring.md — creating components
• stage-6-packaging.md — plugin structure
• stage-7-audit.md — validation
• repo-types.md — CLI vs Library vs MCP patterns

Skills loaded:

• research — external discovery methodology
• codebase-analysis — repo analysis approach
• find-patterns — pattern extraction
• codify — pattern-to-component mapping
• claude-plugins — plugin packaging