Kg RAG Development Design Harness — AI skill for Claude Code

AOSE Harness

An addon layer for Claude Code that turns app and website work into **checkable artifacts**: a plan whose claims are verified by gates that actually run, a design plane grounded in a knowledge graph rather than invented per session, and cold dispatch that can send one frozen specification to several agents and compare what comes back.

It does not write code better than Claude does. It makes what Claude produces **consistent, cited, and provable** — across sessions, surfaces, and agents.

npm test                 # 171 harness tests, 38 example tests
npm run lint             # LINT-01..35 over a blueprint
npm run design:check     # the design gate: 8 checks, measured
npm run kg:ask -- directions cond_audience_solo_operator cond_content_tabular_numeric

---

The idea in one line

**Write down what must be true, and build the thing that checks it. Never one without the other.**

That sounds obvious. It isn't, because writing rules down is easy and checking them gets skipped — and **a rule nobody checks looks exactly like a rule somebody checks.**

This project found five of those in one session across three codebases. The clearest: a design contract declared seven colour-contrast targets on the day the project was initialised, and in all that time **not one had ever been measured**. It read as a guarantee. It did nothing.

See [`docs/declarations-clarification-correction.md`](docs/declarations-clarification-correction.md) for the full catalogue and the rule that follows.

Four planes, one machine

Each plane runs the same five moves: declare the non-negotiables → lint the declaration → a human gate that freezes what it approved → cold execution against the frozen artifact → a scored review with cited evidence.

Plane	Declares	Proves
build	constitution.yaml, specs, tasks	runs the execution gate, captures exit code and stdout hash
look	design.system.yaml	measures contrast, tokens, stat