Forward-deployed / Learning zone
Technical product managementa standalone module
Lesson 02

Discovery to delivery

TL;DR

Products don't move in a straight line from idea to launch. They move through two parallel tracks. Discovery answers "should we build this, and what exactly?" — cheaply, with prototypes, data, and user conversations. Delivery answers "build it well" — with specs, sprints, and releases. The classic failure is skipping discovery and using the delivery track, the expensive one, to find out an idea was wrong. The discipline is dual-track: discovery runs continuously, a step ahead of delivery, and feeds it only ideas that have already survived a risk check — valuable, usable, feasible, viable. Telemetry from shipped work then flows back to seed the next round of discovery.

🎯 For the AI PM

Why it matters — For an AI feature, feasibility can't be assessed in a meeting. It has to be demonstrated. Whether the model can actually do the task, at acceptable quality and cost, is unknowable until you've run it against real examples.

What it changes in your decisions — Discovery for AI features always includes a feasibility spike: a prompt or fine-tune tested against a few dozen real inputs, scored, before the feature enters the delivery track. "The demo looked great" is not a discovery output. A scored eval set is.

Ask yourself — "What evidence do I have that this is valuable, usable, feasible, and viable — and which of those four am I taking on faith?"

Risk if ignored — The team spends a quarter productionizing a feature whose quality ceiling was discoverable in a week, or ships the demo-day version that collapses on real user inputs.

The two tracks

The two tracks

Discovery is cheap to be wrong · delivery is expensive to be wrong

The tracks run concurrently, not sequentially. Release feeds discovery — that's the loop.

Discovery — cheap to be wrong
Opportunity
→
Assess — valuable? viable?
→
Prototype & test — usable? feasible?
→
Kill or reshape ↻
Delivery — expensive to be wrong
Spec & slice
→
Build & verify
→
Release & measure

The tracks run concurrently, not sequentially. While engineering delivers this month's validated bet, you (with a designer and an engineer, part-time) are de-risking next month's. Delivery should never have to stop and wait for discovery to figure out what's next — and discovery should never hand delivery an unvalidated guess.

Discovery: kill ideas cheaply

Marty Cagan's four risks are the standard checklist, and they map to who you need:

The output of discovery is not a document. It's a decision — build, kill, or reshape — plus the evidence that justifies it. Kills are wins: every idea killed for $5k of discovery is a quarter of delivery capacity saved.

Delivery: slice small, verify constantly

Once a bet enters delivery, the craft becomes decomposition. The unit of progress is a vertical slice — a thin, end-to-end piece a user could actually touch (one happy path, one input type) — not a horizontal layer ("first we build the whole backend"). Vertical slices surface integration problems early, produce demo-able progress every sprint, and give you honest options: if the quarter goes sideways, you have a smaller shippable thing, not 80% of everything and 100% of nothing.

Your delivery-track responsibilities, sprint by sprint: keep acceptance criteria unambiguous (next lesson), make scope calls fast when reality diverges from the spec, and protect the definition of done — including instrumentation, so release actually closes the loop back into discovery.

The loop is the point

The diagram's most important arrow is the last one: release feeds discovery. A shipped feature is a hypothesis put into production. Telemetry and user feedback are the results coming back. Teams that skip this arrow ship feature after feature into the void and call it velocity. Put a date in the calendar the day you launch: 2–4 weeks out, review the metrics against what the PRD predicted, and write down what you learned. That document is where next quarter's best discovery leads come from.

Failure modes

Practitioner checklist