The AI Brain of Zo.E: How Gemini Actually Works Inside WF2

Most people assume the interesting part of an AI automation is the workflow plumbing. It isn't. In Zo.E, the plumbing took a weekend. The prompts took weeks.

Here's what actually happens when Zo.E generates content. WF2 is the content engine. It pulls a topic and brand context from NocoDB, assembles a prompt, sends it to Google Gemini Flash, and expects back a single block of structured JSON. That JSON maps directly to NocoDB fields — body_linkedin, body_bluesky, hashtags, image_prompt. If Gemini returns anything other than clean, correctly structured JSON, WF2 breaks. Full stop.

That constraint is the thing that changes how you have to write prompts.

When you chat with an AI, you're having a conversation. Loose language is fine. Partial answers are fine. The model can clarify, hedge, go long. When you're writing prompts for automation, you're writing software. The prompt is a function signature. Gemini is the function. The output has to be deterministic — same shape, every single time, regardless of topic. One stray sentence outside the JSON object, one missing field, one hallucinated extra key, and the downstream nodes fail silently or loudly.

Gemini Flash handles this well. It's fast, it's cheap, and — critically — it's good at structured output. That's why it's the right model for this use case. You're not asking it to reason through a complex problem. You're asking it to take a brief and render it into a specific format, repeatedly, at volume.

The dual-brand capability is where it gets interesting. Nunlimited and SSC have completely different tones — Nunlimited is reflective and founder-focused, SSC is direct and sales-practitioner focused. Same Gemini model. Same WF2 workflow. Different system prompts. The brand voice is injected at the prompt level, not hardcoded into the workflow. That means swapping or tuning a voice is a prompt edit, not a workflow rebuild.

Which brings me to the thing that actually took the most effort: the prompts themselves.

The GEMINI_PROMPTS.md document in your Template Pack is the intellectual property of this system. The workflow architecture is a vehicle. The prompts are the engine. Before you customise anything — before you change a field name, adjust a tone, or add a third brand — read that document first. It explains the prompt structure, why certain instructions are written the way they are, and what breaks if you change them carelessly.

Prompt engineering for automation is not the same skill as prompt engineering for chat. It took longer than I expected to get it right. That document exists so you don't have to start from scratch.

Ta,

James
Founder | Nunlimited