Prompt Helper

The name is not a lie, which is rarer than it should be in this ecosystem. You type "a young woman in a cafe, rainy window, late afternoon" into PromptHelper and it hands back a prompt that's actually shaped for the model you're about to run - a FLUX Kontext edit instruction, a long Z-Image description, comma-separated SDXL phrases, or a booru tag list with score tags for Pony. It's the "LLM writes your prompt" pattern done properly, and it's the node I'd reach for first out of the comfyui-prompt-tools pack.

Why it works so well: if the checkpoint's own encoder is a language model (Qwen3 on Z-Image, T5/instruction models on FLUX Kontext), then having a second LLM write that instruction is just translating between two things that speak the same language. The blank-page problem - sitting there with an empty prompt box - is the thing this kills.

How it works

There's no bundled model. PromptHelper is a thin client: it takes your short prompt, loads a mode-specific system prompt, sends both to an LLM, strips the "Sure, here is your enhanced prompt:" noise, and returns the cleaned rewrite. The engine selector is shared across the whole pack and offers five backends - Ollama (local, no key), vLLM (local, OpenAI-compatible), OpenAI, Claude via OpenRouter (marked experimental by the author, untested), and Gemini. Default is Ollama with qwen3-vl:8b, which is a vision model doing text duty; it works fine, though a plain llama3.2:3b or gemma2:2b is lighter if you're on low VRAM.

The mode dropdown is the whole product. Eleven modes, each with a hand-written system prompt encoding real output rules. FLUX Kontext (Scene Edit) forces action verbs, "ensure correct human anatomy", and "preserve the same facial features" - the exact scaffolding people forget. LTX-2.3 Video (Audio-Video) demands explicit motion and camera behavior plus a synchronized audio line, because LTX ignores negation and invents camera moves otherwise. SDXL Pony/Illustrious switches to booru-style tags. The two Random Character modes pull from configurable pools (ethnicity_pool, age_pool, mood_pool, hair_pool, plus optional lighting, setting, outfit_style) and force a fresh pick on every queue - the node returns NaN from IS_CHANGED so ComfyUI can't cache the result. Nice touch for bulk character ideation.

The inputs that matter

• prompt - your rough idea, the one thing you actually write.
• mode - the model family you're targeting. Pick this before the engine.
• engine / base_url / model - must match each other. The dropdowns are a combined list across all engines (ComfyUI can't filter them), so picking ollama with an OpenAI URL is on you, and it surfaces as a connection error at runtime.
• temperature (default 0.7) - lower it if the output wanders off into invented detail.
• custom_system_prompt - override the mode's system prompt entirely, for the tinkerers.

Two outputs: enhanced_prompt (the rewrite) and original_prompt (your input echoed back, handy for comparing or logging). Wire enhanced_prompt into your CLIPTextEncode positive - or into a TextMux if you want an AI/manual override switch in the same pack.

Installing

Via ComfyUI Manager, search comfyui-prompt-tools (display name "Prompt Tools"). Or manually:

cd /path/to/ComfyUI/custom_nodes
git clone [email protected]:wernerberHH/comfyui-prompt-tools.git
pip install pyyaml   # optional - enables the URL/model autocomplete dropdowns

Then restart. The one dependency is pyyaml, and it's genuinely optional - without it the URL and model fields degrade to plain text inputs and the nodes still work. No model downloads: the LLM lives in your Ollama/vLLM server (ollama pull qwen3-vl:8b) or behind an API key you enter in Settings → PromptTools.

Common issues

Engine/URL/model mismatch is the #1 runtime error - match the three fields. If ComfyUI runs in Docker with default bridge networking, it can't reach your Ollama on the host as localhost; use network_mode: host or host.docker.internal (the README calls this out explicitly). Expect subject drift from small models - an 8B enhancer adds detail you didn't ask for, and chaining narrowly-scoped steps drifts less than one free rewrite. And the usual caveat for any LLM node: it's arbitrary Python with network access, so read what a fresh pack does. This one is MIT-licensed and keeps API keys in a 0600 gitignored file that never leaves your host, which is more than most.