APNext H3 Claude Code Continue Writer

MiniMax-H3 renders clips, not movies. Every video node takes a prompt and produces one chunk - 5 to 15 seconds, give or take - and if you want a longer piece you have to keep the story going across clips. APNext H3 Claude Code Continue Writer is the node that does the hand-off: it looks at the frames of the clip you just rendered, shows them to Claude Code, and writes the H3 prompt for the next clip - with the last frame already designated as the new clip's first frame.

That last bit is the whole trick. Video models are allergic to continuity you don't spell out, so this node writes an I2VA prompt (image-to-video: the first frame is a given image) where the last frame of the old clip is <Picture 1> at 0.00s of the new clip. It hands that frame back out as first_frame so you can wire it straight into the video node. Continuation becomes a matter of "start here, go there" instead of "hopefully start near here."

How it samples what Claude sees

You feed the rendered clip's frames into frames (a decoded video batch). The node doesn't dump the whole thing at the model - it takes frame_count frames every frame_stride frames, counting back from the end. Defaults are 4 frames, one every 6 (a quarter second at 24fps), and the final frame is always included. The earlier frames are context only: they tell the model which way the camera and the action were moving, so the continuation doesn't reverse a pan or freeze mid-swing. Those sampled frames come back out as context_frames so you can preview exactly what it saw - worth doing once, because if the sampling strip looks wrong, the continuation will too.

Switch continuation_mode to T2VA and it instead writes a cut to a fresh scene of the same story; the frames only inform look and mood.

Inputs that matter

• frames - the rendered clip (decode it first).
• frame_count / frame_stride - how many frames Claude sees and how far apart.
• idea - what happens next. Leave empty and Claude decides how the scene naturally carries on; that's often the better first pass.
• previous_prompt - the H3 prompt that made the clip you're continuing. For a long chain, paste every prompt so far, oldest first. Not needed if you use resume_session_id.
• resume_session_id - the killer feature. Feed it the writer's session_id (or the previous Continue Writer's) and Claude Code still has the guide, the earlier frames and every prompt in context, so characters, speaker IDs and the soundscape stay consistent across many clips.
• duration_seconds, visual_style (Auto keeps the previous clip's look - almost always what you want), wildness, the camera block, dialogue toggles - same controls as the base writer.
• llm (optional) - connect an H3 LLM Backend to write with Ollama instead of the CLI.

Outputs: h3_prompt plus the split-out section fields (integrated_multimodal_description, overall_soundscape, non_diegetic_music), session_id, info, first_frame and context_frames.

Installing

Part of dagthomas/comfyui_dagthomas. ComfyUI Manager → search comfyui_dagthomas, or clone into custom_nodes + pip install -r requirements.txt, restart. Under APNext → H3.

The one real prerequisite: Claude Code must be installed and logged in (claude on PATH, or CLAUDE_CODE_PATH set). If claude isn't found, the node can't run - there's no API-key fallback for the default engine. Calls are slower than an API hit (tens of seconds; 25–60s is normal for an H3 prompt, longer with research on), and each call carries Claude Code's ~35k-token tool preamble, so a long chain burns your subscription window - that's the trade for a session that remembers everything. Keep timeout_seconds above the default 900 for research runs, and don't forget use_subscription is on by default, which is what makes it bill your seat rather than an API key.