# Lab #02 — Playbook v1 (Baseline)
**Date locked:** 2026-04-27
**Status:** Active baseline
**Hypothesis:** None (this is the reference implementation against which v2+ will measure).
## Configuration snapshot
| Dimension | v1 setting |
|---|---|
| Channels | Claude Code (primary), Codex (mirror — declarative parity) |
| Phases | 10 (1 / 1.5 / 2 / 3 / 4 / 5 / 6 / 6.5 / 7 / 8 + handoff) |
| Sub-agents | 8 (master, ux-researcher, asset-curator, ui-junior, microcopy, a11y-auditor, persona-tester, critic) |
| Iteration cap | 3 |
| User checkpoints | 3 mandatory (Phase 3 / Phase 5 / Phase 8) + 1 informational (Phase 1.5 asset choice) |
| Discovery questions | 5-8 in one batch |
| UX research parallelism | 2-3 ux-researcher agents |
| Persona testing parallelism | 4 persona-testers (1 from §13 user-defined + 3 adversarial) |
| Adversarial persona budget | 3s/step, ABANDON token strict |
| Eval pipeline | deterministic gate → prompt-diversity Claude jury → persona walkthrough |
| External API keys | NOT required — orchestration runs entirely inside Claude Code / Codex session |
| Cross-family jury | Opt-in v2+ feature — v1 uses 3-judge prompt-diversity ensemble on host model |
| Vision-LLM use | Triage only with Set-of-Mark |
| Self-Refine cap | 2 retries gated on oracle signal |
| Korean persona source | Nemotron-Personas-Korea lazy-load opt-in (PIPA-compliant) |
| Asset fallback chain | Pinterest URL listing → Unsplash → Pexels → Lucide → Heroicons → unDraw → Mixkit → Coverr → brand-color SVG |
| Asset Pinterest policy | URL listing only (TOS forbids scraping) |
| Run archive | Permanent, content-addressed screenshots |
| Token budget | None (per user: explicitly out of scope for v1) |
## Agent prompts source
The 8 prompts live at `.claude/agents/omd-*.md` (this repo) and `.codex/agents/omd-*.toml` (Codex mirror). v1 ships them verbatim.
## Eval rubric (jury)
LLM jury rates against 4 dimensions (0-5 each, no compound score):
| Dimension | Rubric criteria |
|---|---|
| Voice fidelity | Forbidden phrases used: 0. Microcopy matches §10 register. |
| State completeness | All 5 states (empty/loading/error/success/skeleton) present per screen. |
| Principle adherence | No §12 Principle violations in wireframes/components. |
| IA discipline | Every screen has entry, exit, error path. |
Pairwise vs previous iteration, both orderings, position bias mitigated.
## Persona pool (default)
For Korean projects:
- 1 from DESIGN.md §13 (user-defined; if missing, swap to nemotron 1×)
- adversarial-impatient (3s budget)
- adversarial-low-literacy
- adversarial-brand-skeptic
For English projects: same skeleton, persona prompts in English.
## Success criteria for v1
A v1 run is "successful" if:
1. Reaches Phase 9 handoff without crash or infinite loop.
2. All 3 mandatory user checkpoints honored (no auto-skip).
3. handoff/v0.zip is non-empty and contains brief + DESIGN.md + wireframes + components/manifest.json + assets/.
4. eval/deterministic.json verdict = "pass".
5. ≥ 2/4 personas reach task_success = true.
6. Total tokens stay under ANY single-day Claude Pro / ChatGPT Plus quota for an active user (informally tracked, not enforced).
## Known v1 limitations (will guide v2-v3 hypotheses)
- Cross-family jury graceful-degrades to single-family if user lacks GPT/Gemini keys → v2 candidate: enforce 2-family minimum or substitute with prompt-diversity ensemble.
- Asset sourcing is API-key-gated (Unsplash / Pexels APIs); without keys, falls back to WebFetch on public pages → v2 candidate: ship with default API keys via package.json env-template, or bundle a public-domain seed pack.
- Persona walkthrough simulation mode (no Playwright) loses motion/click realism → v2 candidate: bundle a static-HTML render layer so simulation has DOM to walk.
- Korean persona pool requires user opt-in (license) → v2 candidate: pre-stage opt-in at first run instead of mid-flow.
## Instrumentation (what v1 measures, mandatory for compare/)
Every v1 run must emit `postmortem.md` with:
```markdown
- iterations: <N>
- total_tokens_estimated: <approx based on agent log>
- persona_abandon_rate: <abandons / total personas>
- deterministic_pass_rate: <passes / total checks>
- jury_score_normalized: <avg rubric score / 5>
- time_to_handoff_min: <ISO end - ISO start, in minutes>
- user_satisfied: yes | partial | no | unknown
- user_reject_phases: [<phases user pushed back on>]
- agent_token_breakdown:
omd-master: ~Xk
omd-ux-researcher (×3): ~Xk
...
- failure_modes_hit: [<list of named failure modes>]
```
`omd lab compare` reads these to build metrics.json.
## Scope freeze
v1 does NOT include:
- Pixel rendering of components (delegated to v0/Cursor/Subframe)
- A11y violations beyond axe-core + lighthouse (no manual heuristic walk)
- Live A/B testing with real users (synthetic personas only)
- Automatic API key acquisition (user provides)
These are explicit scope decisions — moving them in is a v2+ change, not a v1 bug.