Braze

Promptfree, for everyone, and the only version there is

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

Copy

Be clear about what you are replacing before you start. **Braze is bought for scale**, and if your problem is orchestrating messages to tens of millions of people across four channels in real time, this is not a weekend project and the honest answer is to keep paying.

If your problem is the same shape at a smaller size — and it usually is — here is the buildable version.

STACK
- Node 20+ with Fastify
- SQLite through better-sqlite3, WAL mode, until volume genuinely demands otherwise
- A worker process, or several
- A relay for email, a push provider for devices
- Caddy in front

THE MODEL
- profiles: id, external_id, attrs_json, channels_json — email, device tokens, phone, each with its own consent and timestamp
- events: id, profile_id, name, props_json, at
- campaigns and canvases: ordered steps with waits, conditions and channel sends
- enrolments with version, position and exit reason
- sends: one row per message per channel, with the provider reference
- experiments: id, name, variants_json, holdout_percent

ORCHESTRATION ACROSS CHANNELS, WHICH IS THE POINT
- **One frequency cap across every channel.** The failure this product prevents is somebody getting an email, a push and an SMS about the same thing in ten minutes, and a homemade system with a cap per channel produces exactly that
- Channel preference per profile, honoured, with a fallback order — push if the app is installed, otherwise email
- Quiet hours in the profile's own timezone, enforced across all channels

PUSH, WHICH HAS ITS OWN RULES
- Rent the delivery; the platform gateways are not buildable
- **Handle token invalidation properly.** Dead tokens accumulate fast, and sending to them wastes quota and eventually gets you throttled — process the feedback and remove them
- A push is an interruption on somebody's personal device. Cap it far more strictly than email, and make the unsubscribe one tap

EXPERIMENTATION, DONE HONESTLY
- Variants assigned deterministically from a hash of the profile and the experiment, so somebody stays in one arm
- **A permanent global holdout** — a small slice that receives nothing from any campaign. It is the only way to measure whether the whole programme is worth running, and it is the number that eventually justifies or kills the budget
- Report with confidence intervals, and refuse to declare a winner below a stated sample size

RELIABILITY
- Queue with leases, idempotency per profile per step, retries with backoff, dead-letter with replay
- Versioned enrolments, so editing a canvas does not move people mid-flight
- **A kill switch per campaign that takes effect in seconds.** At this scale a mistake reaches a great many people before anybody can react, and the ability to stop it is worth more than any feature

WHAT MATTERS MOST
The cross-channel frequency cap, the global holdout and the kill switch. The first protects the customer, the second proves the value, the third limits the damage.