APNext H3 Claude Code Reference Writer

This is the Claude Code version of the H3 Reference Prompt Writer: same six-section full-reference format, same <Subject N>/<Picture N> label rules, same retention_analysis - but the model writing it is your locally installed Claude Code CLI, billing your subscription seat instead of an API key. If you've read the plain H3 Ref Prompt Writer article, you know the format; this one just answers the "who writes it and what can it look up" question differently.

It inherits every option from the reference writer - idea, task_type (the bracketed summary prefix: keyframe completion, reference generation, video editing, video continuation, audio reuse, audio reference), reference_role, word_target, shot_plan, visual_style, wildness, the camera vocabulary, include_dialogue and dialogue_language, plus four reference image inputs (image_1–image_4) and reference_notes for describing what must survive or for video/audio references you can't attach.

Why this variant

Two reasons, same as its base-format sibling. The research toggle sends the CLI to the web before writing - the real location, the wardrobe, the light, the physics - and folds what it finds into the prompt as concrete visual detail, with no citations and no commentary in the output. For full-reference work that's genuinely useful: reference images give you the subject, research gives you the world it moves through, and both end up in the same structured prompt.

And it returns a session_id instead of model_used. That's the hook for refinement - wire it into the H3 Claude Code Refiner and describe the change in plain language ("keep the cardigan, move the market to dawn") with the guide, your reference images and the model's reasoning still in context. It also accepts a resume_session_id of its own and an optional working_dir for a script or lookbook the CLI may read.

Model aliases (sonnet, opus, haiku, fable, default), use_subscription, timeout_seconds (default 900 - reference rewrites are long) and seed all behave as on the other Claude Code nodes.

Install

Pack install is the shared routine - 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

Restart, then make sure the Claude Code CLI is installed and logged in once (claude on PATH, or CLAUDE_CODE_PATH set). No ComfyUI-side API key needed. The reference images and guide files live in the pack already - nothing extra to download.

Common issues

• References not sticking - the model can only track what it can see; verify your images are actually connected to image_1–image_4, and lean on reference_notes to say plainly what must be preserved.
• Slow with research on - a full-reference rewrite plus web fetches is a long agent loop. Bump timeout_seconds rather than killing the queue mid-run.
• Subscription window - heavy batches can exhaust your plan's rolling limit; the node prints a warning when it hits it. That's the price of the seat.