LLenny's Podcast
← All frameworks
ProductivityNick Turley (Head of ChatGPT at OpenAI)

Two-Mode Prioritization for AI Products

Prioritize backward from model magic AND forward from customer needs

Difficulty
Advanced
Time to result
~months to results
Steps
5
Confidence
88%

Building on top of a rapidly improving model requires two distinct prioritization modes running in parallel. Mode one works backward from a new model capability — find the magic thing the tech can now do and make it shine (art, not a PM framework). Mode two is classic product management — listen to customers, find the primitives they share across segments, and build those. Applying a standard PM framework to mode-one work will lead you badly astray.

Origin

Turley describes how he prioritizes across ChatGPT's consumer, enterprise, and general surfaces. He explicitly warns that if you applied a rigid PM framework to a breakthrough capability like GPT-5's front-end coding, 'you would do something horribly wrong.' Voice was a mode-one win: not customer-requested, but a magic capability they found a creative way to productize.

Core principles

  • 01Model capabilities and customer needs demand different prioritization logic
  • 02When a capability jumps, re-plan and re-prioritize around it — don't force it through segmentation
  • 03Working backward from the model is art, not science
  • 04Across very different customer segments, look for overlapping primitives before building segment-specific features
  • 05Litmus test: a good feature should get ~2x better as the model gets 2x smarter

How to run it

  1. 1

    Inventory the model magic

    Ask what breakthrough capability the tech now has — e.g. GPT-5 became genuinely good at front-end coding.

    Pro tip Voice and coding weren't customer requests — they were 'wow, we figured out how' moments worth productizing creatively.

  2. 2

    Re-plan around the capability (mode one)

    Reprioritize to bring the capability to life in the most awesome way. Do NOT run it through a standard PM framework — that's how you get it horribly wrong.

    Pro tip This is more important than any particular audience segmentation for genuinely novel capabilities.

  3. 3

    Find the shared primitives (mode two)

    For customer-driven work, look across segments (home, work, school, enterprise) for the large overlap — projects, history, search, sharing, collaboration — and invest there for maximum mileage.

    Pro tip General-purpose products have far more cross-segment overlap than you'd expect; exploit it before building bespoke.

  4. 4

    Do the non-negotiable segment work

    Handle the genuinely segment-specific requirements separately — HIPAA, SOC 2, deployment/privacy for enterprise are non-negotiable if you want to be a serious player.

    Watch out Don't let mode-one excitement crowd out the unglamorous compliance work that gates whole markets.

  5. 5

    Apply the 2x litmus test

    For each capability feature, ask: does it get ~2x better as the model gets 2x smarter? If not, it may not be worth shipping.

    Pro tip Compliance features (SOC 2) don't scale with model intelligence — that's fine; the test is for core capabilities.

In the wild

Voice as a capability-first win

Voice wasn't driven by customers begging for it. OpenAI found a way to make the models handle anything-in, anything-out, then asked what a creative way to productize that would be, and shipped to see what people did.

A retentive capability that emerged from working backward from model magic, not from a customer request.

Cross-segment primitives

Turley found that projects, history, search, sharing, and collaboration are wanted almost identically whether users are at work, home, or school — so they became high-mileage shared investments rather than per-segment builds.

Small team ships broadly useful primitives instead of fragmenting effort across segments.

Common mistakes

Forcing a breakthrough capability through a standard PM framework

Turley: 'if you applied some sort of PM framework to that I think you would do something horribly wrong.' Novel capabilities need art and re-planning, not scorecards.

Building segment-specific features before checking for overlap

Different segments look different but share most primitives. Missing the overlap wastes a small team's throughput.

Is it for you?

Best for

PMs building on top of a fast-improving foundation model, juggling multiple very different customer segments

Not ideal for

Mature products with stable capabilities and a single well-understood customer, where classic prioritization suffices

From the transcript

One is sort of working backwards from the model capabilities and that is much more art than science

44:30

if you applied to some sort of PM framework to that I think you would do something horribly wrong

45:00

the other chunk of it really is more like classic product management where you need to listen to customers

46:00

if you know we're shipping a feature and it doesn't get 2x better as the model gets 2x smarter, it's probably not a feature we…

1:04:30

From the episode

Inside ChatGPT: The fastest-growing product in history

Nick Turley (Head of ChatGPT at OpenAI)