APNext H3 Presentation Writer

The name undersells it. You don't write a presentation with this node - you hand it a pile of source material (findings, benchmark numbers, a changelog, release notes, code) and it directs a video of someone presenting it: a presenter on camera walks through the material scene by scene, charts and numbers on screen, script, visuals and pacing all generated. No audio input. The talk is the output.

This is one of the pack's MiniMax-H3 prompt writers, so know the division of labour. H3 is MiniMax's 33B omni-modal video model: it renders 4–15 second clips with native stereo audio, and its weights opened in August 2026 (mind the licence's fine print on who may run them). This node doesn't render anything. It writes the prompt - one complete, spec-compliant H3 prompt per scene. You feed scenes, durations and lengths into a MiniMax H3 video node, then stitch the clips with H3 Scenes Join, and that's your talk.

How it actually works

The writing is done by your locally installed Claude Code CLI, not an API call. That's the pack's whole trick: it uses Claude Code's own login, so the work counts against your Claude subscription seat rather than metered API spend, and Claude Code can be agentic about it. Under the hood the node loads the official MiniMax writing guides as the system prompt and adds a contract it calls "FACTS ARE SACRED" - every number, unit, percentage, date and proper noun spoken or shown comes verbatim from your source_material. It never rounds, extrapolates or "improves" a value. That's the point: the chart values are true to your data, because an unconstrained LLM will happily invent a nicer number.

The material is split into scene_count scenes (default 6, up to 24), each a full four-section H3 prompt with its own duration. Scene 01 hooks and names the topic, the middle scenes take one point each, the last lands the takeaway. The model writes in chunks (scenes_per_call, default 4) so a run stays inside the watchdog; a timed-out chunk is retried at half size. A 6-scene run with research on can take 10–20 minutes, so timeout_seconds (1200 by default) is worth knowing about.

The inputs that matter

• source_material - your ground truth. Paste everything the talk should contain.
• direction - who presents, where, and the tone; empty means the model stages it to fit the material.
• presentation_format - how it's staged: keynote stage with a giant LED screen, whiteboard explainer, news studio, lab demo, boardroom pitch, documentary, or tech screencast.
• scene_count and visual_aids - how many scenes, and how often a chart is on screen (Auto / Every scene / Key data moments / None).

Everything else - visual_style (1,100+ curated looks), dialogue_language, wardrobe/location locks, reference images - behaves like the rest of the pack's writers. One deliberate quirk: wildness here only loosens the staging (weird camera moves, surreal set dressing). The facts, chart values and on-screen text stay verbatim at every level.

Installing it

Install the pack the normal way - ComfyUI Manager, search "comfyui_dagthomas", or:

cd ComfyUI/custom_nodes
git clone https://github.com/dagthomas/comfyui_dagthomas
cd comfyui_dagthomas
pip install -r requirements.txt

Then restart. The pack's requirements.txt is just SDKs - no model downloads for this node. The real dependency is Claude Code itself: install the CLI, run it once to log in, and make sure claude is on PATH. No API key lives in ComfyUI.

Where people get burned

First, it's your subscription seat, and Claude Code's window is a rolling 5-hour rate limit (rateLimitType: five_hour). Heavy multi-scene runs can exhaust it; the node prints a clear warning. The cost=$… figure in info is the CLI's estimate of equivalent API cost, not a charge. Second, research (off by default) sends Claude Code to the web first - real location, wardrobe, light - which is genuinely better but slow and reaches the internet. Third, the director skills cost a few extra seconds and tokens per run; they're the reason the output follows the H3 spec instead of a paraphrase, so leave them on. And if you want a cheap local rewrite instead of the CLI, connect an APNext H3 LLM Backend to the llm socket - the node then writes with Ollama or any OpenAI-compatible server.

Wire scenes (list), durations (seconds per scene) and lengths (H3 frame counts) to the video node; script and synopsis are for reference, session_id chains a refiner. Attached images come back out on image_1..9, so the numbering the prompt sees and the video node sees never drift.