APNext H3 Claude Code Scenes Writer

Writing one H3 prompt with APNext H3 Claude Code Writer gets you a clip. Writing a story means writing a run of clips that hang together - and that's the whole pitch of APNext H3 Claude Code Scenes Writer: one idea in, 1–10 consecutive T2VA prompts out, each a complete base-format H3 prompt (the integrated_multimodal_description / overall_soundscape / non_diegetic_music three-field contract) with its own duration, forming one continuous story.

The Claude Code CLI writes them through the same director skills, camera vocabulary, dialogue toggles and wildness bands as the base writer - the official MiniMax guides in data/h3/ sit in the system prompt, so the model follows the real spec instead of a paraphrase. Each scene is written with a hand-off to the next, and the outputs are lists: scenes and durations have one element per scene, which is what lets a downstream video node render the whole run.

Inputs that matter

• idea - the story, sequence, or idea to expand. Say what happens; the node handles the H3 grammar.
• scene_count - 1–10. Each becomes one element of the scenes list.
• duration_mode - Fixed (every scene = scene_duration) or Vary 5–15s, where Claude paces each scene to fit the beat of the story.
• continuity_mode - the important one. Independent clips writes each scene as its own T2V clip with hard cuts and "already speaking" openers. Continuous chain writes for C2V / motion-context chaining: scene N+1 opens on scene N's last frame, one continuous take, a 2-second silent hand-off before a new speaker's first line. Pick this if you plan to render via Contex Loop or motion-context.
• visual_style - kept identical across the whole run so scene 7 doesn't suddenly switch film stocks.
• wildness, shot_plan, the camera block, dialogue toggles - as on the single writer.
• Optional image - context only; the scenes are T2VA, so nothing becomes <Picture N>. The typed subject_1..3 / scenery_1..3 / object_1..3 sockets are the better tool here: they describe who's who, where it is, and what matters into every scene in words, so the characters and the location stay consistent without the video model ever seeing the pictures.
• wardrobe / locations - the continuity locks, one line per character/place, copied word-for-word into every shot. enforce_wardrobe then parses the output and fires one repair turn if anything got dropped.

Outputs: scenes (list), durations (list), scenes_text (all scenes in === SCENE NN | duration: S.S === envelopes, for previewing), synopsis, scene_count, session_id and info.

Where the lists go

Wire scenes into the H3 video node's prompt and it runs once per scene - a full queue renders the whole story. For iteration, drop H3 Scene Pick in front to render one scene; for one output file, finish with H3 Scenes Join; for true scene-to-scene continuity, route through H3 Scenes → Contex Loop Plan. The pack's h3_scenes_batch.json and h3_scenes_pick_one.json examples show both.

Installing

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

The prerequisite is Claude Code itself - installed, logged in, on PATH. Each scene costs a real Claude Code turn, so a 10-scene run with the director skills on is minutes and a chunk of your subscription window; for longer jobs the pack writes in chunks that continue one session. Don't want to pay in seats at all? Connect an H3 LLM Backend to the llm socket and the same run happens on Ollama - research is ignored there, but everything else behaves.