APNext H3 Short Film Writer

What it actually does

Drop a story, a treatment, a script or even a synopsis into manuscript, and this node hands you back a whole short film's worth of MiniMax H3 prompts - one complete scene prompt per clip, ready for the render side. It's the film-adaptation sibling of the same pack's Music Video Writer: that node paces a video by the song, this one paces it by the story.

Be clear about what it isn't: it doesn't render anything. It writes the prompts; ComfyUI's MiniMax H3 Reference to Video node turns them into actual clips, so you need the H3 weights on disk before any of this becomes a movie. What you get out is the whole plan - scenes, durations, a dialogue script, score notes - written in one coherent session so the film holds together instead of feeling like eight disconnected one-liners.

How the writing works

The manuscript is the contract. Named characters, events, places and any written dialogue survive verbatim into the scenes; the model only invents connective tissue where the manuscript is silent. Then:

1. Size it. length_mode = Scene count gives you exactly scene_count scenes (1–80). Target length takes target_minutes and derives the count at roughly 11 s per scene, the model pacing each clip to land near it.
2. Plan, then chunk. A first call fixes a synopsis with a full Beats: plan plus wardrobe/location locks. Scenes are then written in chunks of scenes_per_call (default 4). With parallel_chunks on, up to 4 chunks at a time are drafted by draft_model (default haiku, several times faster than sonnet) and one continuity pass repairs drift; turn it off for the classic serial run that keeps one long session.
3. Craft baked in. Three-act turns at roughly ¼ and ¾, coverage variety, motivated light, and a score holding one musical voice across the film. wildness is a pure scale here - 0 = sober, 100 = unhinged staging - but unlike other writers it never injects surreal events, so facts and dialogue stay put.

The inputs that matter

• manuscript - your source text. The whole point.
• length_mode + scene_count or target_minutes - how big the film is.
• wildness - 0 to 100, the one dial most people actually touch.
• continuity_mode - Independent clips (hard cuts, each scene its own T2V opener) vs Continuous chain (written for C2V / motion-context chaining so scene N+1 opens on scene N's last frames).
• model - sonnet default, with opus / haiku / fable / default aliases, codex when the Codex CLI is installed, and any ollama: / lmstudio: / local: models your local servers were serving at page load. Or plug an APNext H3 LLM Backend into the llm socket to run offline.
• wardrobe / locations - optional lock lines (one per character or place). Empty = the model fixes its own in the synopsis; enforce_wardrobe (on by default) triggers one repair turn if any shot drops an anchor.

The backend: your Claude Code subscription, not an API key

It's called Claude Code because that's the default engine. The node shells out to the local CLI and uses its login, so with use_subscription on (default) your ANTHROPIC_API_KEY is hidden and the work counts against your subscription seat, not API billing. No API key field anywhere. research sends the agent to the web for real locations and period wardrobe; director (on) loads the pack's H3 skill library. Don't want to touch the seat? A local model through the llm socket works too.

What comes out

scenes, durations and lengths are matching lists - wire them straight into the H3 video node's prompt and length. scenes_text is the whole film for previewing, synopsis the story summary, and script the dialogue written out as a script. session_id lets you continue the conversation in a refiner node later; image_1..9 pass your reference pictures through so the prompt numbering and the video node's sockets can't drift apart.

Installing it

It ships in dagthomas/comfyui_dagthomas ("SDXL Auto Prompter"). Easiest: ComfyUI Manager → search comfyui_dagthomas → Install, then restart. Or:

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

The node downloads no models itself, but you need two external things: the Claude Code CLI installed and logged in once (claude on PATH, or point CLAUDE_CODE_PATH at it), and the MiniMax H3 weights for rendering - a hefty download, and one the H3 Community License restricts in the US, EU, UK and Korea, so check that before you commit.

Where people get burned

• It's slow on purpose. It's a full agent loop, not an API call: 25–60 s per H3 prompt, 10–20 minutes for a multi-scene chunk with director and research on. timeout_seconds defaults to 1200 as the node's own watchdog; a timed-out chunk retries at half size automatically.
• Your subscription has a rolling window. Heavy batches can exhaust the five-hour limit; the node prints a clear warning.
• No temperature, no max_tokens. The CLI owns sampling, so vary the prompt to vary the output.
• Start small. The default of eight scenes is already a real film; don't set scene_count to 80 on your first run.
• On a very old copy, a 2024-era bug changed the working directory on load and broke ComfyUI's Restart button - long fixed, keep the pack updated.