The PM cuts a third of my request, then three specialists build
The product-manager agent reads the research first and writes the PRD. Then a product designer, a prototyper, and a tech architect work from that PRD in parallel. This is the stage where my original ask got edited, on the record.
What this stage wrote to disk
4 agents · 4 documents + the prototypeThe scope decision
I asked for three things. One got killed.Reorder
Extends the app's existing "pin what you use to the top" idea (Starred) from content to navigation.
Hide, reversibly
Hidden items drop into a "More" group, one tap away. Never destructive, and "New chat" can never be hidden.
Default landing screen
No positive comparator precedent anywhere. Changes app-open behavior for a majority well served by chat home. Triggered a kill condition in the competitive research.
prd.md · product-manager"The idea as proposed is three features: reorder + hide + set-default-landing-section. The research does not support shipping all three, and it does not support treating this as an obvious win."
"Two hard gates precede heavy investment… Gate 1 (skewed-usage cohort) — Validation Experiment 1 must show the target cohort exists at meaningful scale before build."
The architect prices it honestly
tech-spec.md · tech-architect"Neither kill condition is cleanly triggered, but the money is in the wrong place if you build UI-first. Build the model and the telemetry first (they are the durable, reusable assets); do not fund per-platform edit-UI parity until Gate 1 confirms the cohort exists."
"Reusing Starred here means reusing a mental model for users, not a code path for engineers. Treat nav preferences as a new subsystem, and price it as one."
Its effort math, carried into the verdict: roughly 1 eng-week for the instrumentation, 7 to 12 eng-weeks for an Android-only beta, 12 to 21 for cross-platform parity. Those numbers are why "instrument first" is a real decision and not hedging.
The prototype, playable
built to the step-2 design languageWorking drag-to-reorder, reversible hide into "More," the locked New-chat anchor, and a first-run discovery prompt. State lives in memory only. Open the hamburger menu, tap Edit, and drag.
The prototyper also filed a manifest separating what the demo proves from what it merely depicts, which is what let the risk sweep audit it honestly:
Real interaction
- Drag-to-reorder, touch and mouse
- Hide / unhide with a "More" group
- Locked "New chat" anchor
- Done / Cancel / Reset
- First-run discovery coach mark
Static or faked
- Status bar, greeting, composer
- No real navigation or routing
- No sync, no persistence
- Almost none of its own accessibility intent