Product Team — Domain Orchestrator & Discovery Loop

This orchestrator does two jobs. Routing: fork context, classify a product inquiry with scripts/product_goal_router.py across all 16 product-team lanes (12 bundled + 4 standalone plugins), run exactly one, return a digest. Looping: run product work as bounded agentic loops with machine-checkable gates — the continuous-discovery loop (weekly cadence scored by discovery_cadence_tracker.py, tree structure enforced by ost_linter.py) and goal-scale runs through the repo-wide agent-harness.

When to invoke

Routing logic (deterministic)

python3 scripts/product_goal_router.py --text "<the goal>" --output json

Exit 0 → route_to names the skill (with skill_path, including the standalone plugins): load its SKILL.md and follow its workflow. Exit 2 → ask ONE clarifying question naming the listed candidates, with a recommended answer. Exit 3 → no signal: ask the user to restate the goal with the deliverable named. Never guess silently; never silently chain — digest first, confirm, then chain.

The discovery loop (the domain's recurring agentic loop)

Modern discovery is a weekly habit, not a project phase (Torres). Run it as a bounded loop with two machine gates:

1. Observe — maintain discovery_log.json (interviews, assumption tests; shape in assets/sample_discovery_log.json) and score the cadence:

   python3 scripts/discovery_cadence_tracker.py --input discovery_log.json
   

   Refuses on < 2 interviews (exit 5) — there is no cadence to measure yet. Output: health 0–100, verdict HEALTHY/AT-RISK/DORMANT, named gaps, and next_loop_action.
2. Choose — the tracker's next_loop_action IS the choice: book the touchpoint, re-anchor the guide on the outcome, or test the top untested assumption (route to product-discovery's assumption_mapper for prioritization).
3. Act — run the interview / assumption test with the routed sub-skill's tools.
4. Verify — keep the tree structurally sound before it may drive a roadmap:

   python3 scripts/ost_linter.py --input ost.json    # exit 2 = NEEDS-REWORK, fix before citing the tree
   

   Rules: one measurable outcome root (O1), opportunities are needs not features (O2), targeted opportunities compare ≥ 2 solutions (O3), every solution has an assumption test (O4), no orphan solutions (O5 — the feature-factory tell).
5. Record / Repeat-or-stop — update the log, keep the weekly streak alive. Stop states: HEALTHY + validated assumption → graduate to experiment-designer (build the A/B gate) or product-manager-toolkit (PRD); DORMANT for 4+ weeks → escalate to the product lead by name — do not quietly let discovery die.

For build-scale goals ("turn this validated spec into a repo and verify it"), compile through the repo-wide harness instead:

python3 engineering/agent-harness/skills/agent-harness/scripts/goal_compiler.py \
  --goal "<goal>" --manifest engineering/agent-harness/skills/agent-harness/assets/harnesses/product-team.json \
  --out .agent-harness/plan.json

The domain's three strongest close-out gates plug in as task verifications: ../spec-to-repo/scripts/validate_project.py (exit 0), code-to-prd's golden expected_outputs/, and research-summarizer's citation-count check.

Hard rules

1. Evidence before conviction: no roadmap item cites the OST unless ost_linter.py exits 0; no insight is asserted from a single participant (anecdote, not insight).
2. Outcome-first: every loop hangs from one measurable outcome — the linter's O1 rule is the intake gate.
3. Experiments are gated by math: sample size from ../experiment-designer/scripts/sample_size_calculator.py, never gut feel; report the MDE with the verdict.
4. Prioritization shows its framework: RICE for steady-state, WSJF/cost-of-delay when time sensitivity dominates, opportunity scoring for underserved needs — name which and why (see references/product_operating_model.md).
5. AI features ship with evals: a golden set + rubric is the PRD's quality contract for probabilistic features (references/ai_product_evals.md).
6. Never modify a gate you are judged by; exhausted budgets escalate to a named human, never report as success.

Forcing-question library (grill-with-docs pattern)

One per turn, recommended answer, canon citation. Never run a sub-skill or start a loop until the lane-defining decision is locked:

• DISCOVERY lane: "What is the single outcome this discovery serves, stated with a number? Recommended: write it as the OST root first — opportunities without an outcome are a feature factory. Canon: Torres, Continuous Discovery Habits; opportunity solution trees (producttalk.org)."
• PRIORITIZE lane: "Does time sensitivity change this ranking — would delaying any item a quarter erode its value? Recommended: if yes, run WSJF/cost-of-delay alongside RICE and compare ranks; flag items whose rank flips on a one-step estimate change. Canon: Reinertsen, Principles of Product Development Flow; SAFe WSJF false-precision critique."
• EXPERIMENT lane: "What baseline rate and MDE justify this test's runtime? Recommended: compute n first; if you can't reach it in 4 weeks, test a bigger lever. Canon: statistical power analysis (experiment-designer)."
• ANALYTICS lane: "Is your North Star a leading indicator of value exchange, or revenue/vanity? Recommended: leading value metric with an input tree. Canon: Amplitude, The North Star Playbook."
• STRATEGY lane: "Are these OKRs outcomes or shipping lists? Recommended: outcomes — output OKRs are the #1 operating-model failure. Canon: Cagan, Transformed (SVPG, 2024)."
• BUILD lanes (spec-to-repo / saas-scaffolder): "Which validated assumption says this should be built at all? Recommended: link the OST test that survived; building is the most expensive way to test an idea. Canon: Torres; Bland, Testing Business Ideas."

Assumptions

1. The user owns (or advises the owner of) the product decision.
2. Discovery data lives in the workspace as JSON logs — the loop is file-backed and resumable; every tool ships --sample so the shape is visible first.
3. The four standalone plugins are installed alongside the bundle (the router still routes to them by path if not).

Non-goals

• Not the delivery loop — sprint/flow/Jira work routes to project-management.
• Not the generic loop engine — that is engineering/agent-harness; this orchestrator is the product-domain adapter (router + discovery gates).
• Not campaign marketing — marketing/landing builds from-scratch marketing pages; landing-page-generator here scaffolds product Next.js/TSX pages.

Output artifacts

Anti-patterns (do not)

• ❌ Run all 16 lanes "to be thorough" — route to one, digest, chain on confirmation
• ❌ Cite an OST that fails the linter, or promote a single-participant anecdote to insight
• ❌ Ship an AI feature whose PRD has no eval (golden set + rubric)
• ❌ Let the discovery streak die silently — DORMANT escalates by name
• ❌ Treat RICE as the only prioritization lens when deadlines dominate

References

• references/continuous_discovery_canon.md — Torres, OST, assumption testing, JTBD switch interviews, story mapping
• references/product_operating_model.md — Cagan Transformed, North Star framework, PLG benchmarks, WSJF/ODI vs RICE
• references/ai_product_evals.md — evals-as-PRD, model cards, evaluator-optimizer loops
• Loop engine: engineering/agent-harness · Loop vocabulary: loop-library