---

name: cleanup description: "Repository cleanup with evidence-based decisions (user)" version: 1.0.0 metadata: priority: medium requires: ["claude-mem"] changelog: "1.0.0 - Initial skill from command conversion"

Repository Cleanup

Perform comprehensive repository cleanup with evidence-based decisions.

When to Activate

• User says "cleanup", "clean up repo", "tidy files"
• Working directory has accumulated unstaged/untracked files

Workflow

Launch the cleanup-analyzer agent to scan unstaged and untracked files, then process categories interactively.

Phase 1: Analysis (Agent)

Launch cleanup-analyzer agent using the Task tool with subagent_type='general-purpose'. The agent will return JSON with categorized files and evidence.

Phase 2: Quick Wins (Interactive)

Present obvious_deletions and markdown files:

• Show evidence for each deletion candidate
• Execute approved deletions immediately
• For .md files not in lessons_learned_docs, offer to run /ingest-docs

Phase 3: Deduplication (Interactive)

For each duplicate pair:

• Show similarity score and evidence
• Show memory context if available
• Present options: merge/keep both/delete one/skip
• Execute approved action immediately

Phase 4: Completed Work (Interactive)

For each completed proposal:

• Show implementation evidence
• Present draft memory (content, tags, files)
• Get approval to store memory + archive file
• Execute: store memory, then git mv to docs/completed/

Phase 5: Should Commit (Interactive)

• Show files ready to commit with evidence
• Get approval and execute git add

Phase 6: Ambiguous (Interactive)

For each ambiguous file:

• Show evidence and questions
• Ask user: keep/delete/commit/defer
• Execute decision

Phase 7: Summary

• Show actions taken (counts by type)
• Draft cleanup session memory
• Store memory if approved

Guardrails

1. Execute actions immediately after approval (don't batch)
2. Show evidence before each decision
3. For .md files, integrate with /ingest-docs workflow
4. Capture completed work and cleanup decisions to memory
5. Use the agent's evidence, don't re-analyze
6. Never delete without explicit approval