documentation-managerlisted
Install: claude install-skill pedroknigge/documentation-manager
# Documentation Manager
Living project knowledge for humans and AI agents. **Code is the source of truth for *how* and for whether a claim is true.**
**AGENTS.md + docs/** capture *what*, *why*, decisions, and plans — but **never override code** when they disagree.
## Core rules
1. Prefer **`AGENTS.md`** as the hub. If `agents.md` already exists, use it. If only `CLAUDE.md` exists, either extend it with a docs index or add `AGENTS.md` and cross-link — do not create three competing hubs.
2. Every generated/updated doc uses relative Markdown links back to the hub and related docs.
3. Never delete durable knowledge without a reason; prefer **Superseded by** notes on ADRs.
4. Do **not** auto-commit or auto-push. Leave git to the user.
5. Write generated content in the **user's language** (or the repo's dominant language). This skill body is English for ecosystem compatibility.
6. Prefer accuracy and usefulness over volume. Core set first; optional docs on demand.
7. **Intent first (project-level).** Classify **Intent**: `integrate` | `audit` | `from-zero` (and optional hybrid). Overlay **`production-harden`** when the user is hardening for production (“no volver a prototipo”). See Step 0. **Not required** for pure feature/plan/sync.
8. **Code wins on conflict.** Doc claims that fail structural verification are marked Contradicted/Missing — do not invent code to match docs; fix or flag the doc.
9. **Integrate-first (when Intent=integrate).** The default tree is a **proposal**