API_PromptHelper☀

Most people write bad image prompts - too vague, too much vibe language, too little structure. API_PromptHelper is the DashuaiTools answer: a text-in, text-out node that sends your raw idea to an LLM along with a long, opinionated system prompt and gets back a prompt that's actually shaped for a text-to-image model. Unlike its sibling API_caption, there's no image involved. This is pure prompt polishing.

What the default instruction actually does

The custom_instruction field ships with a genuinely well-written system prompt, and it's worth reading before you replace it. It tells the model to:

• Preserve the immutable core - subject, quantity, action, state, any named IP, colors, text.
• Reason out a concrete visual solution when the request isn't a direct scene description.
• Add composition, lighting, material texture, color scheme, and depth.
• Transcribe any text that should appear in the image inside double quotes, describing its font and layout.
• Never use meta-labels like "8K" or "masterpiece," and output only the final prompt.

That's the difference between this node and a generic "write me a prompt" call - the default instruction is doing prompt-engineering work you'd otherwise have to argue with a model about.

Inputs

• api_type / api_url / API_Key / model_name - the same provider plumbing as API_caption (Siliconflow, T8zhenzhen, OpenRouter, or Other). Default model here is Pro/moonshotai/Kimi-K2.6, a text LLM.
• custom_instruction - the system prompt; replace it with your own if you want a different flavor.
• prompt - your raw input.
• output_language - Chinese or English.
• thinking_mode, temperature, max_tokens, noise_seed - the shared knobs. Same retry-and-strip-<think> handling as the other API nodes.

Output: text, your polished prompt.

When to use it

This is the node for "I have a vibe but no prompt." It's also handy for the iterative loop where you take a generated image's caption, feed it back in, and refine toward a specific style. The one thing it won't do is fix bad model behavior - a polished prompt still gets fed to whatever sampler and checkpoint you're using.

Gotchas

• Same paid-API caveat as its sibling. Every run costs tokens; the 258 max_tokens default keeps each call cheap.
• The default instruction is English, aimed at a specific aesthetic. It's tuned for "faithful to intent, no metaphor, no filler" output. If you want tag-style or Danbooru output, write your own custom_instruction - the default will fight you.
• Key hygiene. The API key rides in the workflow JSON.

Install

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

then restart ComfyUI, or grab it via ComfyUI Manager. Needs requests, declared by the pack.

If you've ever stared at a blank prompt box, this is the node that turns "something moody with neon" into a structured, usable prompt.