Raw idea → build-ready decision brief.
After installing product-engineering, upstream intent work runs through discovery-loop: frame → explore → converge → commit. The discovery-lead agent diverges across candidate product shapes, drives the lens roster to convergence, and emits a connected hypothesis with validation hooks — no engine. You review at four human gates; the agent runs everything between them.
discovery-loop4 gatesThe discovery supervisor. Diverges across candidate product shapes, drives the lens roster to convergence, and emits a connected hypothesis with validation hooks.
frame-intent1 gateEstablishes product framing — problem, user, outcome — before the loop begins.
frame-domain1 gateGrounds the product in the real-world activity it serves and bounds the MVP — produces Domain Framing and Scope Boundary artifacts before the convergent design loop.
frame-situation1 gateClassifies an initiative-level signal into a typed finding, assesses Wardley capability maturities, and anchors the team to the right entry point in the PE six-step shaping sequence.
identify-opportunities1 gateStep 2 of the PE six-step shaping sequence — surfaces all functional, emotional, and social jobs behind an opportunity area, scores each via the Ulwick formula, and produces a ranked opportunity-assessment.md artifact.
diverge-solutions1 gateGenerates ≥3 structured comparable solution options for a confirmed opportunity, with a recommendation and retained rationale for parked and rejected options.
lean-canvas1 gateElicits an initiative brief through an adapted Lean Canvas (simple 5-box or full 9-box) and produces a single shareable initiative brief with a Value Proposition section.
de-risk-intent1 gateSurfaces the riskiest assumption and designs a prototype approach to de-risk it before committing to build.
place-bet1 gateStep 5 of the PE six-step shaping sequence — commits the team to a chosen direction by producing a structured bet.md with a full betting table for map-capabilities to reason against.
map-capabilities1 gateStep 6 (terminal) of the PE six-step shaping sequence — translates a committed bet into a Capability Map with L1 domain organisation, Wardley × strategic-criticality annotation, Build/Buy/Partner/Adopt disposition, and a Build-only suggested build sequence anchoring M3–M6 spec-writing.
decompose-intent1 gateDecomposes the chosen direction into briefs and specs for the delivery loop.
explore-optionsGenerates candidate product shapes with distinct tradeoff profiles.
plan-validationValidates a plan against the discovery sidecar — checks that the build plan addresses the grounded hypothesis.
voice-and-microcopy1 gateCharacterizes a product's voice and writes blame-free, actionable UI copy.
align-value-stream1 gateCoordinates a multi-component product intent across a business-unit value stream.
You describe a product idea or problem to the discovery-lead agent. The agent activates discovery-loop, runs frame-intent to establish product framing (problem, user, outcome), and emits an intent document.
You: Read the intent document — the framing is only a page. Give concrete corrections if the problem statement is too vague (“users want to collaborate faster” → “the PM can’t see which stories are blocked without checking three tools”). Give G0 consent once the framing is specific enough to eliminate candidates from. This gate sets the direction for everything that follows.
The agent runs explore-options to generate candidate product shapes with distinct tradeoff profiles. Each candidate represents a meaningfully different approach to the problem.
You: Watch as candidates appear. If a candidate is obviously out-of-scope or repeats a prior approach, say so before the lens roster runs — it saves a review cycle. Otherwise, let the diverge step complete.
After generating candidates, the agent runs two discovery reviewers — threat and reliability — against each candidate. Each reviewer reads cold and returns findings. Candidates that fail the reviewers are eliminated.
You: Monitor the review output. If a candidate you want to keep is eliminated, read the finding that killed it — sometimes the finding is correct and you need to hear it; sometimes it rests on a false premise you can correct with one sentence.
The agent surfaces the surviving candidates for a mid-course human check. You review which candidates remain and whether the field feels right.
You: At G1.5, make a call. If only one candidate survives and it feels too easy, ask the agent to re-explore from a different angle. If the survivors feel genuinely distinct, confirm and let the loop converge. This is the last moment to expand the option space cheaply.
The agent runs de-risk-intent to surface the riskiest assumption and design a prototype approach to test it. It then runs decompose-intent to decompose the chosen direction into specs and briefs for the delivery loop.
You: Watch the assumption-test take shape. If the prototype approach the agent proposes is too expensive or won’t actually test the assumption, redirect. A bad de-risk approach means the brief reaches the delivery loop with an untested assumption at its core.
The agent presents the full discovery sidecar — intent, assumption-test, validated candidate, decomposition. You review the G2 reconciliation record and confirm the decision brief is build-ready.
You: Read the full brief — all sections. At G2 you ratify the reconciliation record; at G3 you commit the brief to the delivery loop. These are two distinct decisions: G2 is “is this brief complete?” and G3 is “am I ready to build this?” A brief can be complete but still wrong.
For each gate, everything you need to make a confident decision.
A framing that names a specific user, a falsifiable problem, and a measurable outcome.
A framing that could apply to any product — 'improve developer productivity' is not a problem statement.
G0 is the framing gate. A bad framing means the entire discovery loop explores the wrong space.
A clear field of two or three differentiated candidates, each with a distinct tradeoff profile.
All candidates converged to the same shape before the lens roster ran — the diverge step didn't diverge.
The mid-discovery check prevents the loop from converging on a bad candidate without the human noticing.
A build-ready decision brief — reviewers clean, falsifiable hypothesis, decomposition that fits the delivery loop.
A brief that passes the gate but skips the riskiest assumption. Or a decomposition too large for one loop.
G2 produces the build commitment artifact. A bad brief means the delivery loop builds the wrong thing faithfully.
You ratify the brief and hand it to the delivery loop.
You ratify with reservations and don't surface them. The delivery loop builds faithfully to a brief you had doubts about.
G3 is the discovery-to-delivery handoff. After this gate, the delivery loop builds.
agentbundle install --pack product-engineering --scope user