Content Repurposer Bot

Context

X (Twitter) and LinkedIn have become content machines. Every day, dozens of incredible tweets about AI and marketing fly through the feed — insights that could become posts, articles, valuable content. But the manual process of reading, selecting, adapting, writing, and formatting took hours.

I thought: if I know automation, why not automate this?

The Problem

How to turn viral tweets into ready-to-publish content for LinkedIn and blog — with no manual effort, no cost, and at scale?

Technical Decisions

• n8n as orchestrator — the most flexible automation tool, visual, open source. Zapier and Make are more limited and paid
• Apify for Twitter scraping — Twitter scraping is complex (anti-bot, rate limits, API changes). Apify has ready-made actors that handle it, with a US$5/month free tier that covers the entire project
• Gemini Flash as the AI generator — best free AI for text. 1500 requests/day on free tier, no credit card required. GPT-4 would be more powerful but costs money
• Supabase as the database — real PostgreSQL, generous free tier, native node in n8n
• Telegram as the delivery channel — bot in 2 minutes, no cost, instant notifications on phone

Entire stack on free tier. Monthly cost: $0.

The Process

The pipeline works in stages:

1. Scraping — Apify collects high-engagement tweets about AI and Marketing
2. Scoring — scoring system with weights: likes ×1, retweets ×2, replies ×3. The more qualitative interaction, the higher the score
3. Storage — tweets saved in Supabase with deduplication
4. Generation — Gemini Flash receives the tweet and generates two outputs: LinkedIn post (professional tone, in Portuguese) and SEO blog article with ~1000 words (title, meta description, headings, keywords)
5. Delivery — everything sent via Telegram with formatting, original tweet link, and engagement score

The biggest challenge wasn’t the AI — it was the scraping. Twitter is increasingly restrictive. I tested 4 different Apify actors before finding one that worked on the free tier. I learned that production automation is 80% dealing with rate limits, timeouts, and unexpected formats.

Another issue: synchronous calls to slow APIs don’t work. I had to migrate to an asynchronous pattern (fire → wait → fetch) to avoid timeouts in n8n.

An insight about Gemini Flash: it spent 95% of tokens “thinking” and cut off the response. Turning off thinking mode (thinkingBudget: 0) solved it. More “intelligence” isn’t always better — sometimes you just want the direct output.

And the JSON parser was the node that caused the most trouble. AI doesn’t return perfect JSON every time — you need robust fallbacks.

The Result

• 5-10 posts generated daily automatically (LinkedIn + full SEO blog article)
• ~2 hours saved per day on research and writing
• 76 tweets collected and ranked by engagement on the first run
• A 2-line tweet becomes a 1000-word article with title, meta description, headings, and keywords
• Monthly cost: $0

Lessons Learned

• Production automation is 80% handling edge cases — rate limits, timeouts, unexpected formats, APIs that change
• Asynchronous pattern (fire → wait → fetch) is essential for slow APIs in orchestrators like n8n
• More AI “intelligence” isn’t always better — turning off Gemini Flash’s thinking mode fixed truncated responses
• JSON parser needs robust fallbacks — AI doesn’t return perfect JSON 100% of the time
• A 100% free stack is possible for real projects — well-combined free tiers cover a lot