APNext H3 Claude Code Writer

The H3 Prompt Writer on a subscription instead of an API key. Where the regular writer picks whatever provider has a key set, this one hands the job to your locally installed Claude Code CLI, which authenticates with its own login - no ANTHROPIC_API_KEY in ComfyUI, and the work counts against your Claude plan's seat rather than metered API spend. If you're already paying for Claude, that's a real saving on a workflow that's going to write a lot of prompts.

It writes the same base format as the APNext H3 Prompt Writer - integrated_multimodal_description, overall_soundscape, non_diegetic_music, shot blocks, <d> dialogue tags - from the same official MiniMax guide shipped in data/h3/. All the shared knobs carry over: idea, task_type (T2VA / I2VA / FL2VA / L2VA), duration_seconds, shot_plan, visual_style, wildness, the full camera_motion/amplitude/speed vocabulary, include_dialogue and dialogue_language. The differences are where the Claude Code layer pays off.

The two things that make it worth choosing

research. Flipping this on sends Claude Code to the web before it writes - the real location, period-correct wardrobe, how the light actually behaves there, how the physical event unfolds - and folds what it finds in as concrete visual detail. It's instructed never to cite anything or add commentary, so the output stays a clean H3 prompt instead of a Wikipedia dump. Slower, and it reaches the internet, but it's the difference between "a marketplace in old Cairo" and a prompt that knows what old Cairo smells like on screen. The claude: dropdown entry on the regular writer is fine for a plain one-shot; you pick this node when you want the research pass.

Sessions. This node returns session_id in place of model_used, and it's the whole point of the chain. Feed that into an H3 Claude Code Refiner and you can revise the prompt in plain language - "make shot 2 wilder, cut the dialogue" - with the guide, the reference images and the model's own reasoning still in context. The writer also takes an optional resume_session_id to continue an earlier run, and working_dir to let the CLI read a script, shot list or lookbook notes while it writes.

Model aliases are sonnet, opus, haiku, fable or default. use_subscription (on by default) hides the API key so your login is used; flip it off to bill the API key instead. timeout_seconds defaults to 900 because H3 prompts take 25–60s and research runs take longer. seed only picks the surreal elements at high wildness and controls ComfyUI caching.

Install

The pack install is standard - 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

But the real prerequisite is the Claude Code CLI installed and logged in (claude on PATH, or CLAUDE_CODE_PATH set). The claudecode: model entries only appear when the CLI is found, and this node needs it to exist at all. No API key needed in ComfyUI either way.

Common issues

• research feels slow - it's doing a full agent loop plus web fetches; that's the feature. Raise timeout_seconds rather than cursing it.
• Rate-limited mid-batch - your plan's rolling window (five_hour limit) can exhaust on heavy batches. The node prints a clear warning.
• Empty local model lists don't matter here - this node is Claude Code-specific; the ollama:/lmstudio: options live on the regular writer.