Build Only at the Magic Intersection
Don't ship what anyone could build off the shelf — build only where model and product uniquely meet.
- Difficulty
- Moderate
- Time to result
- ~months to results
- Steps
- 4
- Confidence
- 88%
For a company with a unique asset (here, model post-training access), the test for what to build is: could anyone build this using our models off the shelf? If yes, don't make it your focus. Reserve your effort for the 'magic intersection' where deep product work and the post-training/fine-tuning process combine to do something no off-the-shelf integration could. This means embedding product people directly in the model-training loop, not just prompting a finished model.
Origin
Mike Krieger, building on a Lenny's Summit panel insight (with OpenAI CPO Kevin Weil, moderated by Sarah Guo) that most leverage came from product people embedded with researchers rather than on the product-experience side alone.
Core principles
- 01If off-the-shelf model users could build it, it's not where you should uniquely play
- 02Real differentiation lives at the intersection of product craft and model post-training
- 03Prompting a finished model 'is just not enough' — you need to be in the fine-tuning loop
- 04The functional unit of work shifts from 'take the model, then build a product' to co-shaping the model and product together
- 05There's still room for design to intervene where the model lacks UI taste
How to run it
- 1
Apply the off-the-shelf test
For any proposed feature, ask whether someone could build it just by calling your models off the shelf. If so, deprioritise it as your unique bet.
Pro tip There's still great stuff to be built off the shelf — just recognise it's not your unique advantage.
- 2
Embed product people with researchers
Place product and design people directly into the post-training / model-skills conversations, not downstream of a finished model.
Pro tip Krieger says the highest-leverage move — more than product-on-experience — was product embedded with research.
Watch out PMs who treat the model as fixed and only prompt it will underperform those who get into the training loop.
- 3
Co-shape model and product in a loop
Get into how a capability should work during post-training, then feed learnings from building back into training, iterating both together.
Pro tip Krieger's memory-feature review worked because product had already been talking to researchers for weeks about memory capabilities.
- 4
Let design fix what the model can't
Where the model lacks UI taste or picks a suboptimal solution, have design intervene rather than shipping the model's raw output.
In the wild
Anthropic paired someone from the 'Claude skills' post-training team with product people to revamp how Artifacts works, rather than just taking the model and prompting it a little.
→ Produced an Artifacts experience with Claude 4 that does 'way better than just... use the model and... prompt it a little bit' — a capability at the model-plus-product intersection.
Common mistakes
Shipping features anyone could build off the shelf
If your differentiated asset is the model, building things replicable by any off-the-shelf model user wastes your unique leverage.
Treating the model as a fixed input to product work
PMs who only prompt a finished model miss the fine-tuning loop where the real, defensible product quality is created.
Is it for you?
Best for
Product leaders inside foundation-model or platform companies deciding where to concentrate unique, defensible product effort.
Not ideal for
Application-layer startups without model-training access, whose leverage is deliberately in building well on top of off-the-shelf models.
From the transcript
“if any if we're shipping things that could have been built by anybody just using our models off the shelf”
“But like where we should play and like what we can do uniquely should be stuff that's really at that like magic intersection between the…”
“we like use the model and we like prompted a little bit like that's just not enough we need to be in that like fine-tuning…”
“the functional unit of work at Enthropic is no longer like take the model and then like go like work with design and product to…”
From the episode
Anthropic’s CPO on what comes next
Mike Krieger (co-founder of Instagram)