Split

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 feature flags with local evaluation and an honest measurement story.

STACK
- Node service, SQLite for definitions, a CDN-cached JSON snapshot for delivery
- Client SDK evaluates locally; no network call per flag check, ever

THE FLAG
- Key, default, and a list of rules: match by attribute, or by percentage bucket
- Bucketing by a murmur hash of key plus subject id, so a subject stays in the same bucket across releases and across services
- Every change appended to an audit table: who, when, before, after

DELIVERY
- The whole ruleset published as one JSON document to object storage on every change
- SDKs poll with an ETag, and fall back to the last good snapshot if the fetch fails
- A flag must resolve to its default when everything is unreachable

EXPOSURE EVENTS
- Log an event the first time a subject is exposed to a treatment, not on every evaluation
- Batch and send asynchronously; a dropped exposure must never block the request

MEASUREMENT, WHERE PEOPLE FOOL THEMSELVES
- Join exposures to your own metrics table by subject id
- Report the difference with a confidence interval, not a p-value, and refuse to declare a winner before a preset sample size
- Write the minimum detectable effect in the README, so nobody reads noise as a result

Include a README on the hashing scheme, the snapshot cadence and the stopping rule.