Design Intelligence

Most AI design tooling is a pile of skills with no opinion about when to use which one. This is the opposite. An orchestrator reads your task and routes it through six stages in order, so reference work happens before generation and verification happens before you call it done. The stages are the product. The skills are swappable.

The six stages

• S0 Strategy: run the words through a weak-versus-strong test before a line of copy is written.
• S1 Reference: pull type, palette, and layout signals from real work before building anything.
• S2 Generate: build with the constraints from S0 and S1 already in hand. Every decision traces to a measured reference value or a named audit checkpoint, written as a ledger before the code.
• S2M Motion: when the deliverable moves, a reel, an explainer, a promo, an animated title, name the job of each motion, write the hook, and build so the render reproduces. Verify by rendering and pulling frames, not by eyeballing a green build.
• S3 Audit: for anything that already exists, diagnose before you touch it.
• S4 Verify: a build is not done because it renders. Post-flight checks, then ship.
• S5 Deliver: the right format, the right aspect, the handover that lets the next edit start warm.

What's inside

Four original skills, nothing vendored:

• orchestrator routes the task and enforces the order.
• evidence-led-frontend builds landing pages and portfolios where every decision has a source.
• taste-arbitrage runs copy and positioning through a five-question test.
• mastery-loop is a session operating system that names the real failure patterns: planning-as-avoidance, waiting for perfect conditions, the silent stall.

Why evidence over preferences

A model's preferences are the statistical centre of everything it has seen, and the centre is the slop. evidence-led-frontend allows exactly two sources for any decision: a measured value pulled off a live site, or a named checkpoint from the MX/BX/AIX framework. Anything else is a default wearing a decision's clothes, and it has to be declared as an assumption out loud. The ledger comes before the code, which is also what makes the output defensible to a client rather than just pretty.

Built for Claude Code. Works in Codex, Cursor, and anything that reads agent skills. MIT licensed.