a builder's codex
codex · operators · Cat Wu · ins_research-preview-trust-cadence

Brand work as "research preview" to compress timelines from quarters to weeks

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 A · TL;DR
Brand work as "research preview" to compress timelines from quarters to weeks

Claim

Label new features explicitly as "research preview" to lower the commitment cost of shipping. The reduced expectation lets the team release in a week or two, gather real usage data, and iterate; the tradeoff (less polish at launch) is offset by a published cadence of follow-up shipments.

Mechanism

Most launch friction is reputational: "if we ship this and it breaks, the brand suffers." Pre-naming the artifact as a preview pre-commits the audience to incremental quality. The team then earns trust via iteration cadence, not initial polish. This unsticks the multi-quarter alignment-then-ship pattern that crushes velocity in older orgs.

Conditions

Holds when:

Fails when:

Evidence

"We try to ship most features in research preview within a week. The way you lose trust around quality is by releasing something early and then nothing happens after."

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

Cat's team timelines collapsed from 6 months to 1 month to 1 week to 1 day. Engineers post in an "evergreen launch room" the moment a feature is dog-fooded; Sarah (Docs), Alex (PMM), and DevRel turn around the marketing announcement the next day.

Signals

Counter-evidence

Jenny Wen (Anthropic Design): non-deterministic AI products cannot be mocked, so research-preview is the only honest mode. For deterministic products with stable APIs, traditional GA discipline still wins because users expect contracts, not iterations. The pattern transfers most cleanly to AI-native and consumer-discovery products and least cleanly to platform / API surfaces with SLAs.

Cross-references

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