The Problem
Generating consistent, high-quality LinkedIn content at scale is hard. The typical approaches fail in different ways:
Static prompts produce generic output. You get technically correct posts, but they lack personality and sound like every other AI-generated LinkedIn post. Users can tell it's automated.
Manual writing maintains voice but doesn't scale. Writing 3-5 posts per week while managing a full-time job isn't sustainable.
Hiring writers solves the time problem but creates a new one: how do you transfer your voice, style, and expertise to someone else? Even with detailed briefs, outsourced content often feels off-brand.
Key Decisions
- Built on n8n workflow automation: n8n allows me to orchestrate complex multi-step workflows—topic generation, draft creation, feedback collection, refinement—without writing extensive custom code.
- Continuous feedback integration: After reviewing each draft, I provide specific feedback. That feedback gets stored and fed into the next generation cycle, creating a learning loop.
- Multi-step content generation: Instead of one prompt, the workflow breaks content creation into stages—topic ideation, outline creation, draft writing, refinement—each with its own AI call and human checkpoint.
- Google Sheets as the control layer: Topics, drafts, feedback, and final posts are tracked in Sheets. This gives visibility into the entire content pipeline and makes it easy to audit what's working.
- Combination of Gemini API and Gemma 3 LLM: Different models for different tasks—Gemini for complex reasoning, Gemma for faster iterations.
What I Learned
Feedback loops transform generic AI output into personalized, improving content. The first drafts the agent produced were mediocre. But after 10-15 cycles of feedback, the quality jumped. The agent learned which angles I prefer, which writing style works, and which topics resonate. This isn't achievable with static prompts—it requires iteration and memory.
Workflow automation tools like n8n democratize AI agent building. You don't need to be a machine learning engineer to build sophisticated AI workflows. n8n's visual interface lets you orchestrate API calls, data transformations, and conditional logic without writing backend code. This matters because it allows PMs to prototype AI products quickly.
Human-in-the-loop isn't a limitation—it's a feature. Fully automated content generation sounds appealing, but maintaining quality and authenticity requires human judgment. The best system isn't one that removes humans entirely; it's one that amplifies human expertise through AI assistance.
Model choice matters less than system design. I experimented with multiple LLMs—Gemini, Gemma, others. The quality difference between models was smaller than the quality difference between good system design and poor system design. Feedback loops, structured workflows, and clear checkpoints mattered more than raw model capability.