Formbricks

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 in-product surveys that replace a hosted Formbricks: a question asked at the right moment, to the right user, without shipping code.

Read this first: the product you are replacing is open source and self-hosting it is a real answer — the plan buys a hosted instance somebody else updates. Build this when you need three or four surveys inside your own product and want the targeting to know who your users are.

STACK
- Node 20+ with Fastify, server-rendered admin
- SQLite through better-sqlite3, WAL mode
- One embeddable widget under 15KB gzipped, from my own domain
- Caddy in front

THE DATA MODEL
- surveys: id, name, kind, status, definition_json, trigger_json, targeting_json, display_json, starts_at, ends_at
- questions: part of the definition — id, kind, prompt, config_json, required, logic_json
- people: id, external_id, email, traits_json, first_seen_at, last_seen_at — identified from my own application
- actions: id, person_id, name, meta_json, at — what a user did, which is what makes targeting meaningful
- displays: id, survey_id, person_id, session_id, shown_at, dismissed_at, completed_at
- responses: id, survey_id, person_id, display_id, started_at, finished_at, is_partial, meta_json
- answers: id, response_id, question_id, value_text, value_num, value_json
- Partial responses are kept. Somebody who answered two of five questions has told you something, and discarding it loses the most honest half of the feedback

KINDS
- Link survey: its own page at a token URL, shareable anywhere
- Website survey: a pop-up on a public site, targeted by page and behaviour
- App survey: inside the product, targeted by who the user is and what they have done. This is the one worth building, and the other two fall out of it

TARGETING, WHICH IS WHAT MAKES IT MORE THAN A FORM
- By trait: plan, signup date, usage level, language, any attribute my application passes in
- By action: has done this, has not done that, has done it N times, did it within the last N days
- By page and by element, for website surveys
- Percentage rollout, so a survey goes to a tenth of the eligible people first
- Rules read as sentences in the admin, and a preview says how many people currently match
- Everything evaluated on the server. A widget that downloads the audience list to decide in the browser has leaked the audience list

WHEN IT APPEARS
- Trigger on a page, on a custom event my code sends, on an element being clicked, after a delay, or on exit intent
- Frequency rules that respect the reader: not within N days of any other survey, not more than once per person per survey unless it recurs, not on the first session, and a global cap
- These rules are the difference between a product that asks and a product that nags. Default them conservatively and say why in the admin

QUESTION KINDS
- Rating, scale with labelled ends, net-promoter, single and multiple choice, ranking, free text, date, file upload, a consent checkbox, and a call to action
- Conditional logic as an explicit expression tree over earlier answers, never evaluated code, with the same evaluator on the server and in the widget
- Question order optionally randomised, with the order recorded per response
- The definition is validated on save: an unreachable question or a reference to a later answer is an error before publishing

THE WIDGET
- One script, one init call with a signed identity token from my backend — never a plain user id, which anyone can type
- Shadow DOM so my styles and its styles cannot reach each other
- Positions: a corner card, a bottom bar, a centred modal, or inline in a container
- Keyboard operable, focus trapped in a modal and returned on close, Escape dismisses, announced politely to a screen reader
- Dismissible always, and a dismissal is honoured — a survey that reappears after being closed is the fastest way to make somebody dislike a product
- Answers submitted per question as they are given, so a partial response survives the tab being closed

RESULTS
- Per question: distribution for choices, an average with the distribution for ratings, and the raw list for free text
- Response rate against displays, completion rate against starts, drop-off per question
- Filter every result by any trait or action, which is how 'what do our paying customers think' stops being a second query
- Export CSV and JSON, complete, including partials, with the traits at the time of answering

OPERATIONS
- .env: DATABASE_PATH, BASE_URL, WIDGET_JWT_SECRET, SESSION_SECRET, HASH_SALT
- Migrations on boot, each once
- The targeting payload endpoint answers from memory with an ETag; it is hit on every page of my product
- Nightly backup off the machine, restore script
- Health endpoint

WHAT MATTERS MOST
Targeting on the server and the frequency rules. Build those two before the question types, then put one survey in front of a real audience for a week. A well-aimed question to two hundred of the right people beats a pop-up shown to everybody, and it costs you far less goodwill.

Give me the repository, the widget, migrations, .env.example, a seed with three surveys, and a README with deploy steps behind Caddy.