product-design-harness

# UX3 Product Design Harness

> UX3 is an Agentic Design Process with Designer-in-the-Loop judgment. It
> reviews User, Evidence, and Business flows through six gates and returns one
> verdict before bounded execution: continue, verify, or stop_reframe.

Canonical website:

- https://cis2042.github.io/product-design-harness/

Expanded machine context and canonical answers:

- llms-full.txt
- answers.md
- .well-known/skill-manifest.json
- .well-known/agent-card.json (compatibility copy of the skill manifest)

Canonical contract (every mode returns this):

- schemas/review-result.schema.json

Entry point (selects Quick Gate, Standard Gate, or UX3 Council):

- skills/product-design-harness/SKILL.md
- prompts/start-review.md

Contract-enforced rules (JSON Schema plus scripts/check_review.py, not prose):

- JSON Schema validates output shape, required fields, and conditional branches
- verdict-specific fields are mutually exclusive; a continue result cannot
  carry a reframed_question
- challenge-round fields: forbidden in independent rounds, required in
  challenge rounds (schemas/reviewer-verdict.schema.json)
- every stop_reframe lane verdict carries a stop_reason_class
- actor_boundary.target_population is the single canonical target population
  source in review results and context packs
- scripts/check_review.py is required for combined verdict, weakest flow, and
  headline evidence tier semantics
- evidence tier bounds the verdict: a t0 headline tier forces stop_reframe,
  and t1 can never return continue (scripts/check_review.py)
- may_do and must_not_do cannot share an entry (scripts/check_review.py)
- a ux3_council review needs at least one challenge with a real objection
  (scripts/check_review.py)
- weakest flow: severity, then lowest lane evidence tier, then fixed
  priority evidence/user/business (docs/CONTRACTS.md)
- evidence receipts require provenance, freshness, counter-signal, and a
  tier t0-t4 (schemas/evidence.schema.json)
- human decisions link both ways via review_id and human_decision_record_id
  with accountable owner and reversal conditions
  (schemas/human-decision.schema.json)

Key prompts:

- prompts/start-review.md
- prompts/mode-selector.md
- prompts/gate-review.md
- prompts/ux3-council.md

Agents and integration:

- agents/registry.json
- adapters/implementation-agent.md
- docs/ADOPTION.md

Docs:

- docs/HARNESS.md
- docs/UX3.md
- docs/OPERATING-PROTOCOL.md
- docs/CONTRACTS.md
- docs/COUNCIL.md
- docs/LIVE-CODING.md

Golden review examples (all validate against the canonical review contract):

- examples/quick-gate-review.json
- examples/standard-gate-review.json
- examples/ux3-council-review.json

Live coding golden chain:

- examples/live-coding-product-brief.json
- examples/live-coding-review.json
- examples/live-coding-context-pack.json

Verify locally (CI runs the full suite on every push):

- python3 -m venv .venv
- .venv/bin/python -m pip install -r requirements-dev.txt
- .venv/bin/python -m unittest discover -s tests
- .venv/bin/python scripts/validate.py
- .venv/bin/python scripts/check_review.py examples/quick-gate-review.json
- .venv/bin/python scripts/check_review.py examples/live-coding-review.json

Operating rules:

- review User Flow, Evidence Flow, and Business Flow
- use evidence tiers to bound action
- do not average reviewer verdicts
- run a challenge round in council mode

UX3 Decision Kernel:

- Validate lane contracts and evidence references
- Reduce to the worst verdict and deterministic weakest flow
- Gate uncertainty, risk, reversibility, and human-owned calls
- Emit one canonical verdict, next action, execution boundary, and stop conditions
- keep human-owned decisions human-owned
- create a context pack only after continue