Load Prompts

Ever run an LLM-assisted prompt workflow and watched your favorite result drift away because the model hands back a slightly different prompt on every run? That's the problem Load Prompts exists to solve. It's the save-and-reload side of a two-pack setup: it pairs with IF_AI_Tools, whose "IF Prompt to Prompt" node turns your plain-English question into a proper positive/negative prompt pair using a local Ollama model or any of a pile of LLM APIs. Load Prompts catches that output as it's generated, writes it to a little JSON file, and lets you pull any saved generation back later by index. Think of it as a tiny prompt library with training wheels.

What it actually does

Four things, straight from the README:

1. Auto-saves every generated prompt triplet - question, positive, negative, plus an index - to a file called ai_prompts.json in the plugin's own root folder.
2. Three save modes: append (add each generation to the growing list), rewrite (replace the whole list with each new generation), and clear (wipe the file). More on the append/rewrite caveat below.
3. load_index retrieves the saved entry at a given index, so a prompt you liked last week is one number away.
4. Manual override fields: you can feed in your own Positive and Negative text that gets prepended to the sequence, with a comma-delimited dedup so you never end up with the same quality tag twice.

That last one is genuinely handy. Your model's quality boilerplate (score_9, score_8_up on Pony, masterpiece, best quality on Illustrious, per the KB's prompt-engineering notes) can live once in the Positive field, and when the LLM also happens to return it, the dupes just vanish. You're not re-typing your house style every generation.

The inputs that matter

Only a few are worth touching:

• mode (the only required input): append, rewrite, or clear. Start with append - it's the least destructive and the default.
• Que / Res / Neg: wire these to the Question / Positive / Negative outputs of IF's Prompt-to-Prompt node. They're the incoming generation you're saving.
• load_index: set this to a saved entry's number to reload it. Defaults to 0.
• Positive / Negative: your manual additions, deduped as described above.

The outputs are Question, Positive, and Negative. That's the whole point: Positive goes into your positive CLIP Text Encode, Negative into the negative one, and the loop is closed. No image processing happens here - this node never touches pixels.

Installing it

The easy way: open ComfyUI Manager, search "ComfyUI-ArteMoon", hit install, restart. The manual way:

cd ComfyUI/custom_nodes
git clone https://github.com/Moooonet/ComfyUI-ArteMoon

then restart ComfyUI. No requirements.txt, no model downloads, no API keys - Load Prompts is pure Python stdlib doing JSON reads and string splits. All the heavy machinery lives next door in IF_AI_Tools, which wants Ollama running (or an LLM API key) and, on some systems, poppler installed for its PDF/OCR features. Don't blame this node when that pile is the thing that's heavy.

Gotchas and troubleshooting

• The append/rewrite modes go quiet when the IF Prompt to Prompt node is bypassed. This is the README's own warning: if you disable or mute the LLM node, append and rewrite stop saving (there's no fresh generation to record). clear keeps working regardless. If your JSON file is mysteriously not growing, check whether you bypassed the LLM.
• The file is not where you're looking. ai_prompts.json lives in ComfyUI/custom_nodes/ComfyUI-ArteMoon/ai_prompts.json, not your output folder. Deleting the node folder deletes your saved prompts with it.
• The GitHub repo currently 404s. As of this writing the Moooonet/ComfyUI-ArteMoon repository returns a "page not found" - the owner appears to have taken it down. If you already have the folder in custom_nodes, nothing breaks; it's just local code that keeps running. If you're installing fresh and the clone fails, you may need to track down a fork or mirror.
• Know what you're signing up for. IF_AI_Tools is itself archived, with prompt generation migrated to the IF_AI_PromptImaGen repo and then a community llm-toolkit fork. A beginner googling Load Prompts in 2026 should know both halves of this workflow are semi-abandoned. It still works, but don't build a production pipeline on it - treat it as a fun, instructive experiment in letting a language model do your prompting.

If that experiment is what you want, the loop is satisfying: type a question, watch the LLM draft a prompt, let Load Prompts remember the good ones, and never chase a vanished generation again.