Build for the Current Model, Prototype for the Next
Elicit max capability from today's model while pre-building the products the next model will unlock
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 4
- Confidence
- 91%
The hardest and rarest AI-product skill is being 'the right amount of AGI-pilled': it's easy to design for a hypothetical super-smart future model, but the real work is eliciting maximum capability from the current model — guiding users onto the golden path around its strengths and patching its weaknesses. In parallel, you deliberately prototype products that don't work yet, so that when a stronger model lands you can swap it in and see if the gap closes. As models improve you also delete the crutches you'd added.
Origin
Cat Wu on the core PM skill for AI-native products at Anthropic.
Core principles
- 01Designing for a hypothetical AGI is easy and low-value; eliciting the current model's max is hard and rare
- 02Guide users onto a golden path that plays to model strengths and hides weaknesses
- 03Prototype ahead of capability so you know exactly what's missing
- 04When a new model ships, remove the prompting crutches it no longer needs
- 05Value skill-sets shift every few months — expect large capability jumps and re-plan
How to run it
- 1
Design the golden path for today's model
Instead of building for a future model that can do everything from a text box, figure out how to elicit maximum capability from the current model — guide users to interact with its strengths and patch its weaknesses.
Watch out Being over-AGI-pilled leads you to build an under-scaffolded product the current model can't actually deliver.
- 2
Prototype products that don't work yet
Build the not-yet-possible product deliberately, so you can see precisely what's missing for it to work. The unfinished prototype is a measuring instrument for the capability gap.
- 3
Swap in each new model and re-test the gap
When a stronger model arrives, drop it into the prototype you've already built and check whether it closes the previously-missing gap — turning a dormant prototype into a shippable product ahead of competitors.
- 4
Remove crutches the smarter model no longer needs
On every model launch, re-read the entire system prompt and, section by section, ask whether the model still needs each reminder. Delete the scaffolding it now handles on its own.
Pro tip The product often gets simpler as the model gets smarter — the model 'eats your harness for breakfast.'
In the wild
Early Claude Code would start a 20-call-site refactor, change five, and stop. Sid on the team added a to-do list tool, mimicking how a human tracks call sites, which got the model to complete all 20. With Opus 4 and later models, the model naturally used a to-do approach without prompting, so the feature became a de-emphasized nice-to-have rather than a necessity.
→ A crutch added for a weaker model was removed as the model improved, simplifying the product.
The team tried building a code-review product several times and shipped only simpler versions (a /code-review command). Only with Opus 4.5/4.6 and Sonnet 4.6 was review good enough to run multiple agents across the whole codebase and synthesize real pre-merge issues.
→ The prototype was ready when the model caught up, so the engineering team now relies on the review passing before merging PRs.
Common mistakes
Building for the AGI that hasn't arrived
Designing a bare text-box product for a hypothetical super-smart model leaves today's users without the guidance and scaffolding they need to get value from the current model.
Keeping stale scaffolding after model upgrades
Prompting crutches added for weaker models make the product needlessly complex once the model can do the job itself. Failing to prune them on each launch bloats the harness.
Is it for you?
Best for
AI product builders deciding how much to scaffold around a rapidly-improving model and how to stay ahead of capability jumps
Not ideal for
Stable, non-AI products where the underlying capability isn't changing every few months
From the transcript
“I think the hard thing is figuring out for the current model how do you elicit the maximum capability?”
“How do you help users go get onto the the golden path? How do you like guide users to interact with the model strengths and…”
“it's pretty important to build products that don't necessarily work yet so that you know, okay, what is missing um for this product to work…”
“we can move remove a lot of prompting interventions every time the model gets smarter. And we actually do this every time we launch a…”
From the episode
How Anthropic’s product team moves faster than anyone else
Cat Wu (Head of Product, Claude Code)