The Hook Factory: 5 Prompt Angles I Automated in n8n (Steal the System)

This is part of my build-in-public series on running LLM workflows in n8n in production. Everything here is from a real template I sell and support — the mistakes are mine, the fixes are tested.

Ask a content creator what eats their writing time and they'll say "the draft." It's not the draft. It's the first two lines.

I tracked it for a month: on a typical 800-word post, I spent 40% of the time on the hook and 60% on everything else. The hook decides whether anyone reads the other 800 words, so the stakes justify the time — but the economics don't. Five hooks a week, twenty minutes of staring per hook, and half of them still underperform.

So I did what I always do: I turned the staring into a system, then I turned the system into an n8n workflow. It now generates hooks for every post I write, and I gave the workflow away for free. Here's the prompt system inside it — steal it.

Every hook that stops a scroll is one of five angles wearing different clothes. The workflow runs one prompt per angle, in parallel, and returns five hooks:

1. Curiosity gap — open a loop, don't close it in the hook.

2. Contrarian take — challenge a default belief your reader holds.

3. Specific number — precision signals "I actually measured this."

4. Micro-story — one scene, one stake, no backstory.

5. Direct callout — name the reader, name their pain.

The examples above are real — they're the opening lines of my last five articles. Two of them outperformed the rest of the series combined on comments.

My first version was one prompt: "write 5 hooks in 5 different styles." It failed in a way I now call angle collapse — all five hooks sounded like the same person clearing their throat five times. Same cadence, same vocabulary, just rearranged.

The fix was structural, not linguistic: five separate LLM calls, each with its own system instruction and its own two few-shot examples in that angle's voice. In n8n that's five parallel branches off one Set node holding the topic, merged afterward by a Code node that assembles strict JSON. No single prompt is asked to be five people at once.

Two details that mattered more than I expected:

Five parallel prompts will still occasionally produce near-duplicates — the curiosity-gap hook and the micro-story hook converging on the same sentence. I added a cheap guard in the merge Code node: reject any hook that shares more than 60% of its content words with another hook in the batch, and re-run that angle once. One re-run, not a loop — if it collides twice, I keep both and let the human pick. The human is still in the loop; this is a factory, not an autopilot.

Five LLM calls, ~400 tokens each, on GPT-4o-mini: under a cent per batch of five hooks. The workflow runs on my laptop's n8n before I write a single word of the article. Twenty minutes of staring, replaced by a 12-second run and a 2-minute pick-the-best. That's the whole ROI story.

The full workflow — five angles, the merge node, the dedup guard, the README — is free. It's the lead magnet for my store: if you run n8n, you can import it in three minutes and never stare at a blank caption box again.

The paid AI Content Repurposing Engine is what happens next: it takes the hook and the finished article and produces the X thread, the LinkedIn post, and the newsletter section. Hook first, repurpose second — same pipeline, two products.

I'm building AI Automation Lab — production-ready n8n workflow templates. The free AI Hook Generator is the workflow from this article, import-ready; my open-source templates live at github.com/zhp910318/n8n-ai-automation-lab.