Skip to content
AI-Native PM
The work

Founding Essay · 03

Shape · Ship · Track

The whole job of an AI-native product manager, made concrete. You shape how a model behaves, you ship it to a human, and you track whether it is holding up. Then you do it again, because a probabilistic system is never finished.

7 min read · The framework that turns a model into a product

In "AI-Native Product Management Is a Discipline," we made the case that this is its own discipline: the work of turning a probabilistic system into a product that reliably solves a real problem. In "The Mind on the Other Side of the Model," we showed who the work is for, a human with real and well-documented limits. This is how the work actually gets done.

A base model is raw capability. It can do almost anything, and left alone it reliably does nothing in particular. Turning it into a product people can depend on is the new job of the product manager. We built a framework to operationalize this new job, and used it to ship FuelTheFam, a family nutrition app at fuelthefam.com. We call it Shape · Ship · Track.

Shape · Ship · TrackA continuous cycle of Shape, Ship, and Track, then Shape again, because a probabilistic system is never finished.A probabilistic modelmade reliable1ShapeDecide how it behaves2ShipPut it behind guardrails3TrackCatch what users won't report
The cycle: shape, ship, track, and shape again.

Shape

Decide how the system should behave, and make it behave that way. This is the new skill. The deliverable is no longer a spec of features, it is a spec of behavior: the model you choose, the system prompt that sets the boundaries, the context you feed in, the rules for the cases that go wrong.

When we built FuelTheFam, the hard part was never the screens. It was deciding what the model does when a photo of a dinner plate is blurry, holds three foods at once, or could be chicken or turkey. That behavior is the product.

Shape, the first move On the left, raw model capability fans out in every direction and reliably does nothing in particular. On the right, the same capability is constrained to one intended path. Shape means writing a spec of behavior, not a list of features. SHAPE THE FIRST MOVE RAW CAPABILITY could do anything, reliably nothing SHAPED BEHAVIOR one intended behavior Shape: you write a spec of behavior, not a list of features.
Shape: raw capability, constrained to one intended behavior.

Ship

Put it in front of a human, behind guardrails, at a cost and speed that work. The failure modes of AI are quiet, so this is where you defend against them.

A nutrition number that is confidently wrong throws no error. It just tells a parent their underfed kid is fine. Shipping FuelTheFam meant guardrails for the outputs that should never appear, and a budget that let a parent log a full day of meals in under three minutes. Shipping is also where you earn trust, by designing for the moments the model's output is uncertain instead of hiding them.

Ship, the second move The model sends two outputs toward a person. A good output passes through a gap in the guardrails. A confidently wrong output is stopped at the guardrail before it reaches the human. SHIP THE SECOND MOVE THE MODEL GUARDRAILS confidently wrong THE HUMAN Ship: the quiet failures stopped at the guardrail, before they reach a person.
Ship: guardrails, evals, and a workable budget.

Track

Find out whether it is actually behaving, and catch what users will never tell you. People do not file bugs against an AI that is quietly wrong, they stop trusting it and leave.

So you track model behavior and drift, the slow ways a product gets worse while every dashboard stays green, and you feed what you learn back into Shape. The cycle never closes, because the model underneath you keeps changing.

Track, the third move A model's behavior starts on target and slowly drifts away from it over time while a dashboard would still look fine. The drift is caught and fed back into Shape, so the loop continues. TRACK THE THIRD MOVE TARGET BEHAVIOR drift caught BACK TO SHAPE TIME Track: the slow drift every dashboard misses, fed back into Shape.
Track: catch the drift users never report.

What you do, and who you have to be

The cycle tells you what to do but not whether what you are doing is any good. Run it with no taste and you will ship fast and ship forgettable. "AI-Native Product Management Is a Discipline" went deep on the judgment AI has not commoditized: taste, sense, strategy, product-market fit, and differentiation, and on why AI raised the price of being weak in it. Shaping, shipping, and tracking are what you do, but that judgment is who you have to be, and this is an invitation to run Shape · Ship · Track rather than just read it. A base model is raw capability; what you make it reliably do is the product, and that working product is the proof.

Sources and further reading

This sits alongside the other Founding Essays, and their sources carry forward. The product judgment behind the fundamentals is in "AI-Native Product Management Is a Discipline." The cognitive science behind Ship and Track is in "The Mind on the Other Side of the Model."

The technical practice each move runs on, model selection, system prompts, retrieval-augmented generation, evals for failures that stay silent, and the compounding reliability of agentic systems, is work we go deep on in the pieces that follow, each with its own citations. FuelTheFam is real and live.