← ClaudeAtlas

power-mapperlisted

Enterprise-scale codebase auditing with hierarchical MapReduce. Handles 1M+ LOC with 100% file coverage verification. Use when you need a complete feature inventory of any codebase without missing a single file, page, API endpoint, or integration.
richyparr/power-mapper · ★ 0 · Data & Documents · score 70
Install: claude install-skill richyparr/power-mapper
<essential_principles> **1. Scripts Enumerate, LLMs Understand** Never rely on an LLM to "find" files. Use bash scripts to programmatically list every file, route, endpoint, and integration. Scripts can't miss files. LLMs can and will. **2. Filesystem Is Shared Memory** Agents never pass context to each other. Every agent reads input from disk and writes output to disk. The orchestrator (you) passes file paths, never source code. **3. Hierarchical Compression** Source code (Tier 2) compresses into file analyses. File analyses (Tier 3) compress into domain summaries. Domain summaries (Tier 4-5) compress into product-level features. Each tier compresses ~4x. No agent ever exceeds 200K tokens of meaningful input. **4. 100% Verifiable Coverage** After every audit, a verification script cross-references inventory against analysis outputs. Every file in the inventory must appear in at least one analysis. Coverage < 100% triggers a re-analysis loop for missed files. **5. The Orchestrator Never Reads Source Code** You (the orchestrator) manage the process. You read inventory summaries, create chunk assignments, spawn agents, and verify coverage. You NEVER read source files directly. This keeps your context window free for coordination. </essential_principles> <architecture> ``` Tier 1: INVENTORY bash scripts → .planning/audit/inventory/ ↓ Tier 2: FILE ANALYSIS parallel agents → .planning/audit/files/chunk-{N}.md ↓ Tier 3: DOMAIN SYNTHESIS parallel agents →