AI Video Prompt Engineer (Pro) v2.4 - Akki

The AI Video Prompt Engineer (Pro) v2.4 is the bridge between the still-image side of the AkkiNodes pipeline and the video side. A great image prompt is not automatically a great video prompt - video models need motion language: camera moves, subject movement, pacing. This node takes the master dossier (the established art direction) plus the details of a single shot and produces a video_prompt written for motion rather than composition.

How it works

The source describes it as a three-stage process that ends with an "AI Editor" stage: the model first drafts a verbose, detailed video prompt, then an editor stage condenses it into a tight, motion-focused final prompt. That condensing step matters - video models (Wan, Hunyuan, LTX and friends) respond better to a focused prompt than a paragraph, and the last stage exists precisely because raw LLM output tends to bloat.

The camera_movement and motion_speed drop-downs are populated from wildcard files in the pack (with "Default (from prompt)" and "Random" options). These are your quick creative levers: pick a specific movement like a dolly-in, or hit Random to get variety across shots in the same scene.

Inputs and outputs

Required: llm_model, master_dossier, shot_details, the two drop-downs, temperature (default 0.8 - deliberately higher here, since this is a creative stage where you want variety), seed, and max_tokens defaulting to a tight 1024. That small token budget is the design: the output is supposed to be a single focused prompt, not an essay.

Outputs:

• video_prompt - the final prompt. Feed it into a text-to-video or image-to-video conditioning path.
• full_llm_process_log - the pipeline trace.

Install and gotchas

Pack install: ComfyUI Manager search "AkkiNodes LLM Suite", or git clone https://github.com/routhakash/AkkiNodes-LLM-Suite-for-ComfyUI into custom_nodes, install.bat as admin, GGUF into ComfyUI/models/llms/, restart.

The gotcha is that this node is only as good as the master_dossier and shot_details you hand it - if those are thin or empty, you'll get a generic "a person walks through a room" prompt that no video model will make anything of. It's also tuned for the Akki pipeline's data shapes, so using it standalone means feeding it well-formed text. And remember: the 1024 token cap is intentional, but if your master_dossier is enormous it can eat the budget before the model gets to write the actual prompt - keep the dossier focused on the shot's assets rather than pasting the whole bible.