a builder's codex
codex · operators · Cat Wu · ins_build-products-that-dont-yet-work

Build products at the edge of what does not yet work

By Cat Wu · Head of Product, Claude Code + Co-work, Anthropic · 2026-04-27 · podcast · How Anthropic's product team moves faster than anyone else — Lenny's Podcast

Tier B · TL;DR
Build products at the edge of what does not yet work

Claim

In a fast-improving model environment, the right place to build is at features that almost work today and will work fully on the next model. When the model lands, you swap it in and the gap closes; competitors who waited are six months behind on the surface they have to learn.

Mechanism

Prototyping is now cheap; the expensive thing is being in the market when capability arrives. Products built strictly for the current model collapse to commoditized surfaces immediately because every team can build them. Products built for the next model carry the team's accumulated taste and harness work into the moment the capability becomes general. The runway investment is the harness, not the model.

Conditions

Holds when:

Fails when:

Evidence

"Build products that don't yet work."

Code review at Anthropic was tried multiple times before Opus 4.5 / 4.6 plus Sonnet 4.6 made multi-agent code review reliable enough to gate merges on. Each attempt informed the harness; the final release rode the new model into production immediately.

— Cat Wu on Lenny's Podcast, 2026-04-27

Signals

Counter-evidence

"Build for the next model" is the easy excuse for shipping nothing. The pattern only works if the team actually ships at each model launch — otherwise it is permanent vaporware. Sherwin Wu's variation ("the models will eat your scaffolding") cuts both ways: don't over-invest in the harness either, since the next model may obviate it entirely.

Cross-references

Open the interactive view → View original source → Markdown source →