design-brief

Use when an interface change needs visual or interaction criteria, or a vague design preference must become a bounded design task.

SKILL.md

GitHub

Design brief

Use this optional route for visual/interaction judgment. API or infrastructure work does not need a design score. Reuse the canonical product brief and delivery slice; preserve their OUT/RQ IDs and specification ownership. For a small brand fix, a concise combined record is enough; keep existing components and navigation.

Read docs/templates/design-brief.md and docs/templates/design-rubric.md. Use docs/examples/design-evaluation/library-rooms.md for original anchored examples, not evidence about the current product.

In the ai-dlc source repository, these templates/examples live under agents/templates/ and agents/examples/. In older projects without copied assets, use the record contract below directly; do not require adoption.

Before generation, produce these linked records under docs/design/<work-id>/ (or the project's approved equivalent):

  1. brief.md: owner/status, source brief and delivery slice, OUT/RQ references, audience/journey, supplied evidence versus hypotheses, scope/non-goals, brand constraints, compatibility/accessibility target, unresolved decisions.
  2. rubric.md: ID/version; each criterion's local ID and RQ reference; required behavior with setup/action/expected/evidence; rated criteria with specific 1/3/4 anchors, task thresholds and evidence methods; states/viewports/fixtures.
  3. Run decisions: generator tool, evaluator/access handoff, candidate ID plus revision or digest, iteration ceiling AND time or token ceiling, budget owner, approval source/status and stop/plateau condition.

Keep proposed choices distinct from reviewed task decisions. Permission to use this workflow is not approval for a task budget or paid generation. Missing material decisions or ceilings leave a draft and a concrete next decision. Use discovery/needs-spec when the outcome or behavior needs clarification.

Proposed starters: 0–4 ratings; minimum 3 for required rated criteria; initial candidate plus two revisions; plateau after two comparable evaluations without criterion improvement or a required defect resolved. Projects review/adapt these before generation. Required behavior must pass and required evidence must exist; optional averages cannot compensate. Unobserved is unverified, never 0 or pass. Novelty means useful fit to the brief; appropriate brand reuse is not a defect.

Once the task contract and budget are settled, hand generation to the user's selected existing tool/skill within its authorization. Preserve each candidate and rubric version. Give design-evaluate the brief, rubric, identified candidate and access instructions, keeping generator self-ratings out of the evidence. A separate session or human can evaluate; otherwise label self-review explicitly. Readable artifacts do not establish any client/browser/delegation capability.