Meta Packager

Use this workflow when the user asks Claude Code to turn repeated recent work into reusable assets such as skills, custom subagents, slash commands, hooks, or automations.

Workflow

1. Confirm the packaging scope: target repository assets by default, user-global assets only when the user asks for them, and no live service changes unless explicitly requested.
2. Collect evidence in this order: recent Claude prompts and sessions, project CLAUDE.md, user ~/.claude/CLAUDE.md, shell history, then existing skills, subagents, hooks, commands, plugins, and automations. Keep collection read-only.
3. Inventory existing reusable assets before proposing anything new. Check repository-local assets first, then user-level assets such as ~/.claude/skills, ~/.agents/skills, ~/.claude/agents, ~/.claude/commands, ~/.claude/plugins, and project .claude/ directories.
4. Group candidate patterns by repeated user intent, not exact prompt wording. Count only eligible occurrences from the collected evidence.
5. For each candidate, record occurrence count, stable inputs, expected outputs, tools or commands used, failure modes, and overlapping existing assets.
6. Apply all creation gates before packaging: at least two occurrences, stable inputs and outputs, material time or error-reduction benefit, no existing asset already covers most of the workflow, and no repository policy conflict.
7. Choose the smallest package type that fits: skill for repeatable procedural work, custom subagent for read-heavy parallel triage with a summary returned to the parent, slash command for a thin manual trigger around an existing workflow, hook for deterministic lifecycle enforcement, automation for scheduled or event-driven behavior with clear triggers.
8. Output a shortlist table and stop for explicit user approval before editing files. If the user has already approved a specific candidate in the current request, proceed only with that candidate.
9. Create or extend only approved high-confidence items. Prefer extending an existing asset over creating a parallel skill, subagent, slash command, hook, or automation.
10. For custom subagents, include an explicit reminder that they require deliberate invocation and do not run automatically.
11. Update README or docs/ when adding or changing reusable harness concepts, then run relevant verification.

Evidence Hints