Prompt Quill Generate

Ever stare at an empty prompt box and type a cute cat? because you genuinely have nothing better? That's the exact failure this node was built for. Prompt Quill Generate sends whatever you type to a local Prompt Quill server - a RAG-driven "prompt engineer" that retrieves similar prompts from a vector store of 3.2+ million real ones and has an LLM rewrite yours into something properly detailed. You get back a full prompt and a negative, both as plain strings, ready to feed into a CLIP Text Encode.

Now the honest part: the name is a lie in the good way - it calls no cloud API and needs no key - but the node is only the client. The heavy lifting lives in the separate Prompt Quill server, and that server is the real install: an LLM plus a multi-million-entry prompt store that happily eats VRAM. The ComfyUI package itself is gloriously thin. One Python file, no requirements.txt, and the only import is requests, which ComfyUI already ships. Install this pack and you've done maybe 1% of the work.

How it works

Mechanically there's almost nothing to it, which is the point. The node POSTs your text as JSON to the server's /get_prompt endpoint (default http://127.0.0.1:64738). Server-side, Prompt Quill looks up the most relevant prompts in its store and has an LLM expand and polish yours, then replies with two fields: prompt and neg_prompt. The node returns them both as strings. If you flip add_negative on, your own negative text gets appended to the server's.

The inputs that matter

There are only four, and you'll touch three of them:

• prompt - your seed idea. The default A cute cat? is the author nodding at you; give it a rough scene and let the server do the expanding.
• add_negative + negative - turn the toggle on and your text in negative gets merged into the server's negative prompt.
• url - leave this alone unless Prompt Quill runs on a different machine or port.

Outputs are Prompt and NegativePrompt, both STRING. Wire Prompt into a CLIP Text Encode's positive input, NegativePrompt into the negative, and carry on.

Install

ComfyUI Manager → Custom Nodes Manager → "Install via Git URL", paste the repo, restart. Or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/osi1880vr/prompt_quill_comfyui

Then restart ComfyUI. No pip step, no model downloads from this pack. Then install and run the Prompt Quill server from osi1880vr/prompt_quill on the same machine (or one ComfyUI can reach over the network), because without it this node is a paperweight.

Troubleshooting

• You see "Something went wrong, did you install Prompt Quill?" - that string is literally what the node returns when the HTTP call fails. Server not running, wrong port, or ComfyUI can't reach it. Check the url field and confirm the server answers on http://127.0.0.1:64738.
• It's slow. The LLM has to spin up and answer; first call after a cold start is the worst. Set expectations accordingly.
• Results look 2024-era. The store is overwhelmingly SD1.5/SDXL-style prompts from when it was built. It shines on CLIP-encoded checkpoints; on newer LLM-encoded models (Flux 2, Z-Image, Anima) the verbose output still helps but the long negatives may do close to nothing - the negative prompt is a CLIP-era tool.
• VRAM. Running the server and ComfyUI on one GPU is doable but wants headroom - people report ~16GB to be comfortable. If you're cramped, run the server on a second box and point url at it.

Also be aware the underlying project went quiet in 2025, so what you see is roughly what you get. For a one-stop "make my prompt not suck" node it still works fine - just know the ceiling is the server's data, not the node.