flow

Decision 0068 · Accepted

The experience skill and the
new-skill test

Flow verifies correctness (review) and liveness (ship), and judges fit and metrics (product). Nothing observes a running product the way its user meets it.

Status
Accepted
Built from:
da88c48

Status: Accepted (2026-11-04). Spec: 068. Builds on 0062 and 0067.

Context

Flow verifies correctness (review) and liveness (ship), and judges fit and metrics (product). Nothing observes a running product the way its user meets it. review reads diffs, design proposes looks, and build only says to verify UI in the running app. Catalog growth needs a record (PRODUCT.md).

Decision

  1. Add experience. Optional and read-only: it walks a user journey through a running product with host-native tools and reports observed friction with reproduction steps, impact, and confidence. It is not a lifecycle stage and no skill requires it.
  2. Ownership. design proposes and records visual direction; experience observes the real journey; review verifies the change against accepted intent and flags material defects in a diff; product judges outcomes and direction; ship owns delivery and liveness.
  3. Leading words. "Walk the experience" (experience), "did it land" (product), "work backwards" (explore). "Map the problem" and "show me why" are not registered.
  4. A new skill must pass six questions: a distinct recognizable job; expertise beyond ordinary reasoning; useful evidence, output, or decision; an existing skill that is insufficient; optional and proportional; benefit worth its instruction and maintenance cost. Otherwise improve an existing skill or document the principle.
  5. Not added as skills: land-it, agentic-inquiry, taste-lab, fresh-eyes, state-space, prove-the-boundary, build-the-factory, architecture-scout. Their ideas live as practices inside explore, to-spec, design, review, and product.

Alternatives

  • Fold the walk into review. Rejected: review is bound to a diff and a head; a journey walk needs no change and no head, and would double review's cost on every build.
  • Fold it into design. Rejected: design proposes and records; it also cannot stay read-only.
  • Fold it into product. Rejected: product reads documents, code, and metrics, and never uses the product.
  • Do nothing. Cost: experiential defects surface only when a person happens to look.

Consequences

  • One more skill (about 400 words) in the catalog, with adapter metadata. No eval scenario yet; the first real walk is the evidence.
  • No dependency, recording tool, or gate. Without a browser it reports the limit and what it could still establish.

Jump to, or try a question