Scene 01Case file
Brand Bible
An AI content operating system: raw thoughts come in from Telegram or the web, and a self-critiquing Gemini pipeline turns them into on-brand posts for X and LinkedIn, then schedules and publishes them.
Overview
Most of what a developer could post dies as a half-thought in a notes app.
Brand Bible captures those thoughts where they happen, in a message to a Telegram bot or the web app, and turns them into posts that sound like the author, not like a template.
The core is a recursive drafting pipeline: analyze the idea, pick an angle, draft, critique the draft against the user's brand rules, and refine. Nudges let the user steer a draft without rewriting it, and an auto-publisher with a daily cron posts scheduled content.
It's a Bun monorepo: a Next.js web app with an Elysia API, a serverless grammY Telegram bot, and shared packages for the AI layer (Vercel AI SDK on Vertex AI / Gemini) and the database (Drizzle on Neon Postgres). Auth is better-auth, uploads go through UploadThing, and it installs as a PWA with a ⌘K command palette. Recent work added multi-user onboarding.
What I built
- 01Telegram bot and web capture for raw ideas
- 02Analyze → Angle → Draft → Critique → Refine pipeline
- 03Nudge-based refinement of drafts
- 04Scheduling with cron auto-publishing
- 05Installable PWA with a ⌘K command palette
- 06Multi-user onboarding
Stack
Next.js · Bun (monorepo) · Elysia · Drizzle · Neon PostgreSQL · Vercel AI SDK · Gemini (Vertex AI) · grammY (Telegram) · better-auth · UploadThing