Power Prompt Builder

This is the one node in the CornmeisterNL PowerPack that actually calls out to the internet and costs you money. Power Prompt Builder takes a base idea plus some instructions and uses the OpenAI Responses API to expand it into a finished image prompt. If you've been writing prompts by hand for the new generation of models - where your prompt is a chat instruction to an LLM-based text encoder rather than a bag of tags - this is a way to hand the busywork to a bigger LLM and keep the direction to yourself.

Why it's useful right now: prompting changed. On Z-Image, Flux, and the other LLM-encoded stacks, structured natural-language prompts beat comma soup, and that's exactly the kind of writing a frontier model is good at. The README pitches it for "Z-Image / cinematic prompt styles," which checks out - descriptive block-style prompting is where those models shine. Give it "moody cyberpunk alley, rain, neon" plus some instructions, and it returns a fleshed-out prompt you then feed to a normal CLIP Text Encode.

How it works

Mechanically it's a thin wrapper over the OpenAI Responses API. It POSTs your base_prompt as the user message and instructions as the system message to api.openai.com/v1/responses, with the selected model, temperature, and max_output_tokens. The key comes from the OPENAI_API_KEY environment variable, or from the optional api_key_override field if you'd rather type it in. If neither exists, it raises an error rather than silently doing nothing.

The force_clean_output toggle (on by default) appends "Return ONLY the final image prompt text. No explanations, no bullet points, no markdown, no code fences" to your instructions - the difference between a usable prompt and a ChatGPT essay. The response text gets extracted and handed to you as a single string.

Inputs that matter

• base_prompt - your rough idea. The seed of what you want.
• instructions - how to behave: style guidance, "add cinematic lighting," "make it more detailed," whatever you'd tell a human assistant.
• model - defaults to gpt-5.1; the dropdown offers the GPT-5.x family (chat-latest, base, pro, mini, nano) plus custom..., in which case custom_model lets you type any model name.
• temperature (0–2, default 0.7) and max_output_tokens (default 900) - the standard LLM knobs. Lower temperature for consistent, predictable rewrites.
• force_clean_output - keep it on unless you actually want prose back.

Output: TEXT (STRING) - straight into a CLIP Text Encode.

Install

It ships in the CornmeisterNL PowerPack:

cd ComfyUI/custom_nodes
git clone https://github.com/Cornmeister/ComfyUI-CornmeisterNL-PowerPack.git

Restart ComfyUI, or use ComfyUI Manager → search "CornmeisterNL PowerPack". The pack itself only needs requests and Pillow, but this node additionally needs an OpenAI account and API key - set it with:

export OPENAI_API_KEY=sk-...

before launching ComfyUI (or use api_key_override on the node).

Where people get burned

• "Missing OpenAI API key" - the most common error, and it's exactly what it says. Set the env var or use api_key_override. This is also the one node in a pack that otherwise promises "never break execution": it's a network call, so it will fail loudly if there's no key, no network, or the API errors.
• It costs money and adds latency. Every run is a paid API call. It's not for batch-queuing hundreds of images with a different prompt each - you'll notice the bill and the wait.
• Empty response - the node raises if the API returns nothing useful; bump max_output_tokens or switch models.
• The generated prompt still needs your taste. A strong base_prompt and good instructions are 90% of the result. If the output is generic, that's the input's fault, not the node's.

The honest take: this is a convenience node for people who like LLM-assisted prompt writing and already pay for an API. If you'd rather not hand your prompt generation to OpenAI, skip it - the rest of the pack works fine without it.