WAN Prompt Builder Trio

Wan 2.2 tops out around 81 native frames - about five seconds - and everyone doing real projects works around it by chaining segments. The catch is that hand-writing three separate prompts for the same scene is how you get three videos of the same T-Rex that look like three different movies. Same subject, different lighting, camera teleporting around, character's mood changing for no reason. This node exists to stop that.

WAN Prompt Builder Trio takes the same inputs as its single-segment sibling and returns three coherent prompts for three chained segments of roughly seven seconds each. The system prompt forces a continuous scene - same environment, same light, only the angle changes - with automatic camera progression of wide, then medium, then close. You get one continuous 20-second-ish take instead of three disconnected clips.

How it works

Same Groq call as the rest of the pack (llama-3.3-70b-versatile, free tier fine), but the prompt asks for three outputs in a strict JSON structure. The rules from the Wan 2.2 playbook are baked in per segment: one action only, no chaining words, sentence two is camera and mood. The action intensity slider (1–5) you set gets applied consistently across all three, which is what keeps the energy level from wandering between cuts.

Three outputs - prompt_1, prompt_2, prompt_3, plus a shared negative - all plain strings. The negative is the pack's sensible default (blurry, static camera, low quality, watermark, text, deformed, artifacts, flickering, overexposed), and you can override it.

The one mechanism worth knowing: if Groq fails or returns anything unparseable, the node falls back to three hardcoded templates - wide establishing shot, medium orbit, close push-in - so the workflow never breaks. That's also the tell: if your three prompts suddenly look mechanical and generic, you're looking at fallbacks, not the model's work.

Wiring it in

Each prompt feeds a separate Wan 2.2 text-to-video segment, then you chain those with your usual extension tooling - last-frame-to-first-frame, VACE, or a video joiner. The node deliberately writes the camera progression for you, because mismatched camera language across segments is the fastest route to a visibly spliced result. You'll want to keep your subject and action reasonably specific; the model gets to invent everything you leave vague, and it will.

Installing

ComfyUI Manager (search "WanPromptBuilder") or:

cd ComfyUI/custom_nodes
git clone https://github.com/jideka/ComfyUI-WanPromptBuilder

Restart ComfyUI; nodes land under WAN > prompting. No requirements.txt, no model downloads, no VRAM cost - the pack's only external dependency is requests plus Pillow and numpy, all already in ComfyUI. The actual prerequisite is a Groq API key from console.groq.com; the free tier is plenty for this. The key field defaults to whatever's in the pack's config.json, but the code never writes to that file, so expect to paste the key in yourself.

Honest notes

This is a small, young, one-person pack - launched in May 2026 with the author openly calling it a work in progress. A couple of rough edges follow from that. The JSON parsing is strict: if the model wraps its answer in markdown code fences or adds a preamble, the parse fails and you get fallbacks. Keeping your inputs short helps. And it's a prompt generator, not a consistency guarantee - character identity across segments still drifts the way it always has with Wan chaining, so if you need a specific face, train a LoRA rather than relying on prompt words. For keeping a scene's look and camera logic coherent across three cuts, though, it does exactly what it says.