AI SMM Automation

The challenge

Consistent social media is a grind: someone has to notice what the team shipped, turn it into a post that sounds on-brand (not corporate filler), adapt it per platform, and publish on schedule. For a busy agency that work quietly never gets done.

What we built

An AI content pipeline that turns internal team updates into polished, brand-aligned posts and publishes them across LinkedIn, Facebook, and Threads.

• Team wins are collected in a private Telegram channel
• A weekly job generates platform-specific drafts
• Drafts arrive in Telegram with one-tap Publish / Edit / Regenerate / Discard
• Approved posts publish automatically with managed OAuth tokens
• Engagement metrics feed back to improve future posts

The role of AI

Each cycle runs a multi-stage Claude pipeline per platform:

1. Curate - pick the most compelling narrative from the week's raw updates
2. Write - draft in the platform's format and the DForce brand voice
3. Critique - check the draft against explicit anti-patterns (fake metrics, "excited to share…", safe-but-boring takes)
4. Refine - rewrite once if the critique found issues

The brand voice lives in a cached system prompt, and top-performing past posts are injected as examples so the system keeps getting sharper over time.

Engineering notes

• The approval step is the product. Fully autonomous posting was the obvious design and the wrong one: one off-brand post published unattended costs more than the whole pipeline saves. A human tap is cheap; a retraction is not.
• The critique stage exists because LLM output is competent and boring. A single generation pass produces text that is grammatically fine and completely forgettable. Checking the draft against named anti-patterns and rewriting once is what moves it from acceptable to publishable.
• Prompt caching makes the cadence affordable. The brand voice is a long system prompt reused on every generation across every platform, so caching it is the difference between a weekly job that pays for itself and one that does not.
• Token management is the unglamorous half. OAuth tokens for LinkedIn and Meta expire and get revoked, and a publishing pipeline that silently stops is worse than no pipeline, so refresh and failure surfacing are handled explicitly.

The result

A repeatable engine that produces 15+ on-brand posts a month from work the team was already doing - with a human approving every post in one tap.

Stack: Supabase Edge Functions (Deno), Claude with prompt caching, pg_cron, LinkedIn / Meta APIs, Telegram Bot API.