GrowthBook

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Build me feature flags and experiments that replace a hosted GrowthBook — and know why this is ALMOST.

Flags are a weekend. **The statistics are not.** Deciding whether a variant is better involves sequential testing, multiple comparisons, and the fact that peeking at a running experiment inflates false positives dramatically. Getting that wrong produces confident wrong answers, which is worse than no experiment at all. GrowthBook is open source, so self-hosting it is a real answer; build only if you want the flag layer to be yours and you will use a published method for the maths.

STACK
- Node 20+ with Fastify for the control plane
- SQLite through better-sqlite3, WAL mode, for flags and experiment definitions
- Your existing warehouse or database for the metrics. Do not build an analytics store
- A tiny client SDK, or an evaluation endpoint
- Caddy in front

THE DATA MODEL
- flags: id, key, description, kind, default_value, is_active, created_at
- rules: id, flag_id, position, condition_json, value, rollout_percent, experiment_id — evaluated in order, first match wins
- experiments: id, key, hypothesis, variants_json, traffic_percent, assignment_attribute, started_at, stopped_at, decision, decision_note
- assignments: id, experiment_id, unit_id, variant, at — append-only, and the record that makes analysis possible
- metrics: id, name, kind, query_sql, unit_column, value_column, is_guardrail, direction — defined once, queried against the warehouse
- results: id, experiment_id, metric_id, computed_at, variant_stats_json, method, decision_hint
- Assignments are the one thing you must record from the first minute. Everything else can be recomputed; this cannot

FLAG EVALUATION
- Deterministic: hash the unit identifier with the flag key and bucket. The same user always gets the same variant, on the server and in the browser, with no coordination
- Evaluate on the server wherever possible. A browser that downloads every rule has downloaded your targeting logic and your audience definitions
- The payload is cached with an ETag and fetched once per session; the SDK evaluates locally from it for speed, and the rules in it contain no personal data
- Kill switch: any flag can be forced to a value immediately, and that path must be the simplest code in the system
- Every evaluation of an experiment flag writes an assignment, once per unit, idempotently

THE STATISTICS, WHICH IS THE PART THAT MATTERS
- Choose one published method and implement it faithfully rather than inventing one. A Bayesian approach with a stated prior, or a sequential frequentist test that is valid under continuous monitoring
- **Do not use a fixed-horizon t-test and then look at it daily.** That combination is the single most common error in this field, and it will tell you a neutral change is a winner about a third of the time
- Compute a minimum sample size before starting, from the baseline rate and the effect you care about, and show it beside the running result
- Guardrail metrics evaluated alongside the primary one, with an automatic stop if one degrades beyond a threshold. That is the feature that makes experiments safe to run without watching
- Report an interval, not a point estimate, and label the method next to every number
- A sample-ratio mismatch check on every experiment: if the assignment split is significantly off, the experiment is broken and the result is meaningless. This catches more real bugs than any other check

DECISIONS
- An experiment ends with a written decision and a reason, recorded. Without that, the same question is asked again in six months
- Rolling out the winner is a flag change, not a code change
- Keep stopped experiments and their assignments forever; that history is where organisational learning lives

OPERATIONS
- .env: DATABASE_PATH, WAREHOUSE_URL, BASE_URL, SESSION_SECRET
- Migrations on boot, each once; nightly backup off the machine
- Health endpoint reporting flag payload age and any experiment with a ratio mismatch

WHAT MATTERS MOST
Assignments recorded from day one, a valid sequential method, and the sample-ratio check. Those three are the difference between an experiment programme and an expensive way to confirm what somebody already believed.