I built a zero-budget AI social publisher with n8n — here's what broke
The promise of "zero-budget AI social content" is incredibly tempting. Imagine an automated system that drafts posts, generates accompanying images, and publishes them across your channels, all without touching your wallet. I decided to turn this dream into reality using a potent combination of n8n for workflow orchestration, Google Gemini for text generation, Pollinations.ai for image creation, and Telegram as my command center. What I didn't expect was the sheer volume of little "breaks" I'd encounter along the way, and the creative solutions needed to fix them.
My goal was simple: send a prompt to a Telegram bot, have Gemini draft a social post, Pollinations.ai generate an image based on that post, and then send the complete package back to me for review (or even publish directly). The "free" aspect meant relying on free tiers and open-source solutions.
The Tools of My (Free) Trade
Let's quickly introduce the key players in this automation saga:
- n8n: My workflow automation backbone. Its visual interface is fantastic for connecting various APIs and services without writing much code. I used the self-hosted desktop app, which is completely free.
- Google Gemini (via API): The brain behind the text generation. I leveraged its powerful capabilities to turn concise prompts into engaging social media captions. The free tier offers generous usage limits, perfect for personal projects.
- Pollinations.ai: An incredible open-source platform for generative art and media. It provides a simple API to generate images from text prompts, allowing me to add a visual element to my posts.
- Telegram: My human interface. A simple bot acts as the input for prompts and the output for reviewing generated content.
My First Wall: The Gemini API Conundrum
My initial excitement quickly hit a snag when integrating Gemini. I expected a clean text output, but often, the API would return extra formatting, markdown fragments, or even conversational filler around the desired content.
The Break: Gemini's responses, while powerful, weren't always ready for direct social media publication. For instance, I might get something like:
🟡markdown
Here's a great tweet idea for you:
"Excited about the future of AI! 🤖 What are your thoughts on its impact on daily life? #AI #FutureTech"
🟡
The Fix: I learned to embrace n8n's "Code" node and "Regex" functionality. After the Gemini API call, I inserted a "Code" node with a simple JavaScript function to aggressively strip out unwanted leading/trailing markdown, quotes, and conversational phrases. A simple regular expression like /^['"\\s](.?)(?:['"\s]*$)?/s helped extract the core message. I also implemented a conditional check to ensure the output length was within typical social media limits, prompting a re-generation if it was too long for platforms like X.
Image Generation Glitches: Taming Pollinations.ai's Output
Next up was Pollinations.ai. The API is straightforward: send a text prompt, get an image URL. However, the initial images weren't always consistent in style or directly aligned with the social post's tone.
The Break:
- Varying Image Quality/Style: Without careful prompt engineering for Pollinations.ai, I'd get wildly different aesthetics.
- Slow Generation: Sometimes, the image generation would take longer than expected, causing timeouts or delays in the workflow.
The Fix:
- Refined Image Prompting: I introduced a secondary prompt engineering step before calling Pollinations.ai. Instead of just passing the social post text, I created a "Gemini-powered image prompt generator" node. This node would take the social post, analyze its theme, and craft a detailed, style-specific prompt for Pollinations.ai (e.g., "digital art, vibrant, tech-focused, abstract representation of AI, future city background"). This ensured more consistent and relevant visuals.
- Asynchronous Handling & Fallbacks: For slower generations, I configured the n8n HTTP Request node for Pollinations.ai with a higher timeout. More importantly, I built in a fallback: if the image generation failed or timed out after multiple retries, a "Code" node would trigger a default image URL or simply send the text post without an image, preventing the entire workflow from breaking.
Bringing it Together: Telegram as the Control Center
The final piece was making Telegram an effective two-way street. Not just for input, but for receiving the generated content in a user-friendly format for review.
The Break: Sending raw text and image URLs back to Telegram wasn't visually appealing or convenient for review. I needed a way to present the content as it would appear on a social platform.
The Fix: After Gemini and Pollinations.ai had done their work, I used n8n's "Telegram Send Message" node with Markdown formatting. I constructed a message template that included the generated text, followed by the image (using the [inline-url] syntax for image embeds in Telegram if applicable, or just the URL if embeds weren't perfect). I also added buttons for "Approve & Publish" or "Discard & Regenerate," which would trigger subsequent n8n workflows via a Telegram webhook, creating a true feedback loop.
Lessons Learned and What's Next
Building this zero-budget AI social content pipeline was an eye-opening experience. It reinforced the idea that while free tools offer immense power, they also demand more hands-on configuration, debugging, and creative workarounds. The "breaks" weren't roadblocks; they were opportunities to dive deeper into how APIs actually work and how to make disparate services play nicely together.
The biggest takeaway? You absolutely can build powerful, automated systems without a huge budget, but be prepared to get your hands dirty with data transformation, error handling, and clever prompt engineering. My next step? Integrating a scheduling feature so approved posts get published automatically at optimal times.
The journey of automating content generation with free tools is less about finding a magic bullet and more about the satisfaction of connecting the dots yourself. What surprising 'breaks' have you encountered in your automation projects?