Changefeed

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 a changelog service that replaces Changefeed: I write entries, it publishes a page, a feed, and a widget in my app with an unread badge.

STACK
- Node 20+ with Fastify, server-rendered HTML
- SQLite through better-sqlite3, WAL mode
- One embeddable widget under 12KB gzipped, from my own domain
- Caddy in front

THE DATA MODEL
- projects: id, slug, name, custom_domain, accent, logo_path, is_public
- entries: id, project_id, title, slug, body_md, kind, labels_json, published_at, updated_at, author, cover_path
- reads: id, entry_id, user_key, read_at — user_key is the signed identifier my app passes in
- subscribers: id, project_id, email, token, confirmed_at, unsubscribed_at, cadence
- deliveries: id, entry_id, subscriber_id, sent_at, opened_at — one row per person per entry, so nobody is emailed twice
- views: id, entry_id, path, referrer_host, ip_hash, bot, created_at
- Kinds: new, improved, fixed. Three, not fifteen — a taxonomy nobody can remember gets used at random

WRITING
- Markdown with a live preview, images uploaded and stored locally with the width and height recorded
- Draft, scheduled and published states; a scheduled entry is published by a worker that is idempotent
- Every published entry has a permanent slug and its own page, because a changelog entry is the thing people link to when they say 'this shipped'
- Entries are edited freely before publishing and revised after, with a revision log — a changelog that quietly rewrites history is a bad joke

THE WIDGET, WHICH IS THE PRODUCT
- A script and one init call, given a signed user key from my backend
- An unread badge counting entries published since that user's last read
- Opening it marks entries read as they scroll past, not all at once on open
- Shadow DOM, my accent, dark and light, keyboard navigable, Escape closes, focus returns where it was
- Loads its payload once per session with an ETag; the badge count comes from a tiny separate endpoint that can be called on every page without cost

THE PAGE AND THE FEED
- A public page at my own domain or a subpath, listing entries newest first, filterable by kind and label
- RSS and JSON Feed, both valid, with full content
- Per-entry pages with Open Graph tags and a generated share image carrying the title and the kind
- Sitemap, canonical URLs, and no client-side framework: the page is HTML and it is fast

EMAIL
- Subscribers with double opt-in and one-click unsubscribe that needs no login
- Immediate, weekly or monthly cadence per subscriber
- The digest is assembled from entries not yet delivered to that address, so a change of cadence never sends the same entry twice
- Plain text alongside HTML, and the HTML is tables and inline styles because that is what mail clients render

MEASURING
- Views per entry, widget opens, badge clicks, entries read
- Email delivered, opened, unsubscribed — open rate labelled as the unreliable number it is
- Bots flagged and excluded from published figures
- No third-party analytics, no tracking pixel from anyone else

OPERATIONS
- .env: DATABASE_PATH, BASE_URL, WIDGET_JWT_SECRET, SMTP_URL, SESSION_SECRET, HASH_SALT
- Migrations on boot, each once
- Custom domain via on-demand TLS in Caddy, with the allowed hosts checked against the projects table
- Nightly backup off the machine, restore script
- Health endpoint

WHAT MATTERS MOST
The badge. Build the read state, the count endpoint and the mark-as-read-on-scroll first, and put the widget in a real app. Everything else here is a blog with an RSS feed; the badge is why anyone pays for it.

Give me the repository, the widget script, migrations, .env.example, a seed project with twenty entries, and a README with deploy steps behind Caddy.