Korva is a scroll of the internet’s best ideas.
Claude now runs growth experiments
The CASH project
The project is called CASH, short for Claude Accelerates Sustainable Hypergrowth. It started only a couple of months ago, because before Opus 4.5 the results weren't there, and with Opus 4.6 it began heading in the right direction. The team splits shipping into four parts: identifying opportunities, building the feature, testing it against the quality and brand bar, and analyzing the data after launch. Each part gets its own score for how well Claude does.
Where it stands today
So far the work is small: copy changes and minor UI tweaks. The win rate is about what a junior PM two or three years in would deliver, and below a senior PM's. People still approve each experiment and check brand fit, though a skill holding brand guidelines with clear dos and don'ts could shrink that review over time.
The piece left out
A fifth piece sits outside the loop: cross-functional stakeholder management. Small experiments can skip it, and larger projects can't.
Why score each stage of shipping separately?
Splitting a process into stages shows where a system fails, where a single end-to-end score only shows that it failed. A weak result overall could come from poor ideas, buggy builds, missed brand standards, or misread data, and each needs a different fix. Machine learning teams evaluate pipelines stage by stage for the same reason. Software teams apply the same logic in funnel analysis, which measures drop-off at each step rather than overall conversion.
How are experiment win rates measured?
A win rate is the share of experiments that beat a control version and move the target metric by a statistically meaningful amount. At mature companies most ideas fail that test. Ron Kohavi, Diane Tang, and Ya Xu describe in their book Trustworthy Online Controlled Experiments that at Microsoft roughly a third of tested ideas improved the metrics they targeted, a third had no effect, and a third made them worse.
Why do brand guidelines suit an AI skill?
A skill, as Anthropic uses the term, is a folder of instructions and reference files that Claude loads only when a task calls for it. Brand guidelines fit the format because they are already written as dos and don'ts, which a model can check a draft against. A reviewer can then spot-check outputs instead of reading every change from scratch. Anthropic introduced Agent Skills in October 2025.
More from Lenny's Podcast: Product | Growth | Career
