GrowthBook

Promptfree, for everyone, and the only version there is

Open the prompt in Claude CodeOpen the prompt in CodexOpen the prompt in Cursor

Copy

Build me an experiment assignment service and a results query. The statistics are the risk, not the code.

STACK
- Node service, SQLite for definitions, a JSON snapshot on a CDN for delivery
- Results computed in your existing warehouse, not in the service

ASSIGNMENT
- Hash of experiment key plus subject id, mapped to a bucket in [0,1)
- Deterministic and stable: the same subject gets the same variant across services and deploys
- A subject already exposed never moves, even if traffic weights change; record the weights in force at exposure

EXPOSURE
- Log one row per subject per experiment on first exposure only
- Batched, asynchronous, and never blocking the request

THE ANALYSIS, WHERE EXPERIMENTS ARE USUALLY GOT WRONG
- Join exposures to metrics by subject id in the warehouse; never copy metrics into the service
- Report the relative difference with a confidence interval, never a bare p-value
- Fix the sample size before starting and refuse to display a result before it is reached
- Implement one method properly — a sequential test or a Bayesian posterior — and document which

GUARDRAILS
- Named metrics that must not degrade; the results page shows them beside the primary metric
- An automatic kill switch if a guardrail crosses a threshold with the interval clear of zero

SRM CHECK
- Test the observed split against the intended split every run; a mismatch invalidates the experiment and must be shown loudly

Include a README on the hashing scheme, the stopping rule and the SRM test.