H3 Skill Promptor (Local GGUF)

MiniMax H3 is picky about its prompts. It's a 33B omni video model that renders dialogue, sound and all, and it behaves dramatically better when you feed it the six-section Ref2VA format (or a proper base three-field prompt) instead of a sentence of vibes. The H3 Skill Promptor is the node that gets you there with zero API keys: a local GGUF vision LLM, served by a llama-server process the node spawns for you, writes the whole official-format prompt on your own GPU.

The clever part is how it writes. Most "AI prompt generator" nodes ship a hand-written system prompt that rots as the target model's spec changes. This one loads the actual h3-prompting skill document - live from ~/.claude/skills/h3-prompting/SKILL.md, with a vendored snapshot as fallback - and sends that as the spec, plus exactly one of MiniMax's own worked examples for the mode you picked. Update the skill file and the node follows. The skill is the system prompt, literally.

What you set

The required inputs that actually matter:

• mode - ref2va writes the six-section reference prompt for the MiniMax-H3-Base-Ref2VA checkpoint. t2va / i2va / fl2va / l2va write the three-field base prompts for a different checkpoint, MiniMax-H3-Base-FL2VA (the checkpoint_hint output tells you which you need).
• creative_brief - what the clip should show, in plain language. This is the only writing you do.
• gguf_model / mmproj - picked from models/LLM. mmproj defaults to auto, pairing by filename prefix; set it to none for a text-only model (image inputs then error).
• reasoning_effort - low / medium / xhigh. Note the tooltip: Qwen3.8's template only accepts these three, there's no off switch, and low is the minimum. Each level grants its own thinking budget on top of max_tokens, so a long think can't starve the prompt text.

Optional inputs worth knowing: reference_images (identity/wardrobe/scene refs for ref2va, first/last frames for the base modes), video_frames (a clip as an image batch - keyframes get sampled, max_frames_to_analyze controls how many), dialogue (exact words for <d> tags, kept verbatim), and source_soundscape / source_music / sound_log - those are where you wire in the H3 Audio Soundscaper companion so the prompt's audio sections match what the source clip actually sounds like.

No silent rewrites

Output comes back through a deterministic validator built from the skill's own review checklist. Violations go back to the model exactly once, with the numbered error list; whatever comes back after that is returned as-is. Nothing is quietly "fixed" by the pipeline - the only edits are transport cleanups it reports (a stripped code fence or leaked <think> block). You get three outputs: h3_prompt, checkpoint_hint, and validation_report so you can see what passed, failed, and got retried.

Install and the real prerequisite

Install the pack once: ComfyUI Manager (search "Trent Nodes"), or comfy node registry-install trentnodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/TrentHunter82/TrentNodes.git
cd TrentNodes
pip install -r requirements.txt

Worth a heads-up from the author himself: he renamed the repo on day one, which left two entries in the registry and made Manager occasionally flag the pack as unsafe. If Manager balks, the git clone path above always works.

The load-bearing prerequisite is not pip: this node talks to an external llama-server binary, so you need a current CUDA build of llama.cpp (≥ b10450 for the qwen35 arch family). Put llama-server on PATH, set LLAMA_SERVER_BIN, or leave it at the default ~/llama.cpp/build/bin/llama-server. Models go in ComfyUI/models/LLM - the tested pair is Qwen3.8-27B-UD-Q4_K_XL.gguf (~16.4 GB) plus its mmproj-F16.gguf (~0.9 GB) from unsloth. Any current vision GGUF llama.cpp serves works; these are just the ones it was built against.

Where people get burned

• Empty prompt at high reasoning_effort. Older builds let xhigh spend the entire token budget thinking, and you got a validation failure with no text. The fix is built in now - thinking gets a separate capped allowance - but if you hit the limit-mid-think error anyway, the message tells you straight: raise max_tokens or drop to low.
• "llama-server was not found." That's the missing CUDA build, not the node. Follow the install hint above.
• VRAM refusal. The node checks free VRAM before spawning and refuses rather than OOMing. free_vram_first unloads ComfyUI's own models first - useful when your card is already holding a diffusion model.
• base_url is the escape hatch: point it at any OpenAI-compatible server (LM Studio, vLLM, a llama-server you started by hand) and the node skips managing one entirely.

That last part is the real selling point. Offline, uncensored, no per-call cost, and the prompt format is enforced by the model's own spec rather than your faith in a chat LLM. For the price of ~17 GB of VRAM it's the most dependable way I've found to feed H3 what it actually wants.