review-ai-skillslisted
Install: claude install-skill TsukumoHQ/WRAI.TH
# review-ai-skills — the instructions-are-executable gate
Everything under `skill/` is load-bearing at runtime. `relay.md` is injected into an agent's context and tells it which MCP tools to call and how; `hooks/*.sh` run on every tool use and session start; `llms.txt` is what an LLM quotes when a developer asks it about wrai.th. A wrong sentence here is a wrong *behavior*, fleet-wide — not a typo. Review truth-against-code first, then hook safety, then the behavioral instructions, then the public claims.
Prose style and markdown lint are not this gate's concern. Every item below is something no linter can verify because it requires reading the doc against the actual code.
## 1. Doc ↔ code truth — drift here breaks live agents
- [ ] **Every tool name, category, and count in `relay.md` / `tools-reference.md` resolves to the actual `toolRegistry()` in `internal/relay/toolset.go`.** The registry is the SSOT (64 entries today); `tools-reference.md`'s header still claims "59 tools" — that is exactly the drift to catch, because an agent told a tool exists that the registry doesn't expose gets `tool not found` and stalls, and a tool that exists but is undocumented never gets called. If the diff adds/renames/removes a tool, the doc's names AND the count must move with it.
- [ ] **The default port and URLs match the binary.** `relay.md`, the `.mcp.json` snippet, and the hooks all say `localhost:8090`; `main.go`'s `startServer` defaults to `8090` and `install.sh` sets `DEFAULT_PORT=