Video Prompt Loader v1.0 - Akki

Video prompts in the AkkiNodes suite are generated by the AI Video Prompt Engineer, saved as versioned text files, and then - here's the thing - re-run. The whole video production loop in this pack is "generate, save, iterate." Video Prompt Loader is the part that reads back the latest prompt for a specific shot so you're always sampling against the newest version, not the first one you ever saved.

If you've ever built a text2video workflow where you manually copy a prompt out of a Text node and paste it into a file, this is the node that makes that step unnecessary. It's the read companion to the save nodes, and like the rest of the File I/O tier in this pack, it's pure deterministic Python - no LLM, no API key, nothing to download.

How it works

Three inputs, and two of them are forced inputs (socket-connected):

• project_path - the Project Director's path, relative to ComfyUI's output/ directory.
• video_prompt_subfolder (default DATA) - where the video prompt files live inside the project.
• shot_name - the name of the shot whose prompt you want, e.g. 7B_0005. This is what distinguishes it from Scene Choreography Loader: you look up by shot name, not scene number.

The node searches output/<project_path>/DATA/ for files matching:

VPrompt_7B_0005_3_2026-08-16.txt

That's VPrompt_ + the shot name + _ + a version number + _ + a date + .txt. It finds every file for that shot, sorts them, and returns the newest one's contents as video_prompt. The shot_name input can be wired from something like Shot Selector's SEL_shot_name, which is how the pipeline stays consistent - the name that selected the shot is the name that fetches its prompt.

Installing

It ships with the AkkiNodes LLM Suite, so:

cd ComfyUI/custom_nodes
git clone https://github.com/routhakash/AkkiNodes-LLM-Suite-for-ComfyUI

Windows: right-click install.bat and Run as administrator once - it installs llama-cpp-python==0.3.9 (cu122 wheel) and openai into your ComfyUI venv. Then put a GGUF model in ComfyUI/models/llms/ and restart. On Linux/macOS, install those two packages into your ComfyUI environment manually; there's no cross-platform installer.

Gotchas

• The shot name is a filename token, so keep it filename-safe. The suite's image namer sanitizes names, but if you type a shot name with a / or * into this node, the regex will fail to match anything and you'll get ERROR: No Video Prompt file found for shot ....
• Empty or error shot names are rejected up front. The node explicitly bails if shot_name is empty or starts with ERROR: - so if a downstream error propagates here, it returns a clean error rather than doing something weird.
• "Latest" is determined by filename sort, not file mtime. That's safe because the version and date are baked into the name, but only if you keep using the save nodes' convention. Hand-renamed files break the lookup.
• It returns ERROR: ... strings in the video_prompt output instead of crashing. When your video sampler suddenly gets prompt text that starts with "ERROR:", the fix is usually upstream (the engineer didn't save a file for that shot), not here.

Like most of this pack, Video Prompt Loader has zero community footprint - the code is the documentation. It's small, it's predictable, and in a looping production workflow it's quietly the difference between "works" and "I keep pasting the wrong prompt."