| name | skillify |
| description | Turn a repository or a stable workflow inside a repository into a callable Agent Skill or Claude Code plugin. Use when the user wants a repo exposed as a slash command, reusable skill, plugin, subagent, or installable marketplace package. Performs workflow discovery, a focused interface interview, progressive-disclosure design, scaffolding, validation, and test prompts. Invoke explicitly as /reeper:skillify. |
| disable-model-invocation | true |
| effort | high |
| argument-hint | <repo-path|repo-url> [desired command] |
Reeper Skillify
Package a stable workflow interface, not a repository dump.
Read:
${CLAUDE_SKILL_DIR}/../../references/packaging.md
${CLAUDE_SKILL_DIR}/../../references/security.md
${CLAUDE_SKILL_DIR}/../../references/artifact-contract.md
1. Analyze before designing
Identify:
- what users repeatedly ask the repository to do
- deterministic operations that belong in scripts
- judgment-heavy operations that belong in instructions
- inputs, outputs, side effects, permissions, and failure modes
- required CLIs, services, MCP servers, environment variables, and credentials
- whether it is primarily reference knowledge, an action workflow, a specialized agent, an MCP integration, or a standalone application
Use Repomix or the fingerprint script for large repositories, but read the actual entry points and docs before deciding on commands.
2. Focused interface interview
Ask exactly one material question at a time. Resolve:
- the primary callable job
- expected arguments
- whether Claude or only the user may invoke it
- read-only versus write/destructive behavior
- permission boundaries and approval gates
- required integrations and fallback behavior
- persistent state and generated artifacts
- installation scope: project skill, personal skill, plugin, or marketplace
- command names and discoverability triggers
- success and failure outputs
Recommend the narrowest interface that satisfies the job.
3. Design
Prefer this progression:
- one action skill
- supporting references and deterministic scripts
- subagents only when context isolation or specialized permissions materially help
- plugin wrapper when distribution or multiple components require it
- marketplace wrapper only when installation from a Git repository is desired
Keep each SKILL.md under 500 lines when practical. Put large references in references/, executable helpers in scripts/, templates/assets in their own directories, and eval prompts in evals/.
4. Scaffold
For a standalone skill:
python3 "${CLAUDE_SKILL_DIR}/../../scripts/scaffold_skill.py" \
--output <output-dir> \
--name <skill-name> \
--description "<what and when>"
For a plugin, create:
<plugin>/
├── .claude-plugin/plugin.json
├── skills/<skill-name>/SKILL.md
├── agents/ # only when justified
├── scripts/
├── references/
└── evals/
For marketplace distribution, add a repository-root .claude-plugin/marketplace.json and keep the plugin self-contained beneath its declared source path.
5. Validate and test
Create at least:
- three should-trigger prompts
- three should-not-trigger prompts
- two happy-path execution prompts
- two ambiguity/conflict prompts
- one denied or unsafe operation prompt
- one missing-dependency prompt
Validate locally when Claude Code supports it:
claude plugin validate <marketplace-or-plugin-path>
Test the exact installation and invocation commands. Report any validation that could not be run.
6. Deliver
Provide:
- generated path
- installation command
- callable command(s)
- required dependencies/integrations
- permission and security notes
- test/eval status
- version bump guidance