IAMCCS Prompter — MiniMax H3 Screenplay

The single biggest practical limit of MiniMax H3 is that its prompt has a hard character cap, and the people getting the best results treat it like a screenplay with separate fields - scene, shot list, acting, dialogue, lighting, camera, sound - rather than one blob of prose. IAMCCS_Prompter is a structured writing desk that makes you work that way: you fill the sections, it assembles them in the order the H3 conditioning actually wants, and hands the result downstream as a clean final prompt plus a cine_linx contract.

The default project_data field is a full JSON template (schema v3) with every section pre-labeled: scene, shot_list, acting, dialogue, light_and_image, camera, production_sound, non_diegetic_music, negatives, reference_use, boundary_frames, and the V2V-specific blocks (v2va_subject_definitions, v2va_source_video_authority, v2va_interval_edits, v2va_sound_policy). You don't write JSON by hand - the node ships a UI that edits these fields, and project_json on the output carries the portable version.

The mode switch that matters

• task_mode - t2va, i2va, fl2va, ref2va, v2va_object_swap, audio_driven. This selects which sections actually get composed, so a v2va_object_swap prompt emphasizes the replacement/source-video contract and an audio_driven one leans on the audio_drive_contract and audio_timed_performance blocks.
• injection_target - global, local_auto, or local_1..3. When a Shotboard is downstream, this decides whether your composed prompt replaces/appends the global prompt or lands in a specific timeline slot.
• writing_mode - manual (you write), guided (the desk enforces the structure), assistant_fill (an optional assistant_draft text fills only the empty boxes).
• merge_policy - replace or append.
• character_budget - it watches your character count and refuses to emit a prompt over H3's absolute limit (the error tells you exactly how many characters to cut - genuinely helpful, not cryptic).

Optional cine_linx in is the interesting one: it lets an upstream IAMCCS Cine H3 Vision Info node route analyzed visual context (what the reference frame actually shows) into the declared global/local target without embedding image tensors into the prompt node itself.

Outputs

final_prompt (the assembled plain-text prompt - the thing you wire into a standalone H3 conditioning node), cine_linx (for the Shotboard path, which also gets a visible INJECT action in the UI), project_json, and report.

The honest take

For the Shotboard workflow this is the right tool and it removes a real class of error - "I wrote the prompt in the wrong mode and the references got treated as style instead of identity." Used standalone it's just a well-structured prompt assembler, and there are cheaper ones if all you need is a text box. The value is the contract: the same structured project can drive a Shotboard injection or a direct native conditioning node, and it keeps your scene data as a reusable file rather than a widget you'll misplace.

Install via ComfyUI Manager (search "IAMCCS") or cd ComfyUI/custom_nodes && git clone https://github.com/IAMCCS/IAMCCS-nodes.git. The pack's own user guide (docs/IAMCCS_PROMPTER_MINIMAX_H3_USER_GUIDE_EN.txt) walks the two connection methods - Shotboard injection vs. direct final_prompt - and it's worth a skim before your first serious run.