Prompt Core

If you test LoRAs for a living, you've typed some version of the same prompt two hundred times: a subject name, an outfit description, a scene, and a pile of style words. SickOllie's Prompt Core (Classic family, class SOPromptLogEngine) is the node that makes you stop doing that. It takes one source prompt plus reusable components - outfits A/B/C, a scene, a NAME, an ITEM - and assembles a final string at execution time, so the same graph can produce a hundred variations by rotating text files instead of editing a text box.

How it works

The engine reads from one of two places: manual_prompt typed directly, or a prompt_log_file picked from the .txt libraries under ComfyUI/input/SickOllieLogs/prompts/. Each component stream - prompt, three outfit streams, and scene - has its own file picker and its own mode dropdown: fixed, increment, decrement, randomize, or shuffle. That's how you build a test campaign without a single extra node: randomize the outfit, increment the scene, keep the prompt fixed.

Tokens do the substitution. The default tokens are NAME, OUTFIT_A/OUTFIT_B/OUTFIT_C, SCENE, and ITEM, each with a configurable token string and a value. NAME typically comes from your Loader Core's clean_name output; the others come from their log files. There's also an optional prefix/suffix with its own separator (default , ) and a cleanup pass that strips regex matches - the default rule removes things like [x], (y), and trailing :weight artifacts, which is exactly the junk logs accumulate.

On each run it resolves the indexes, applies the tokens, cleans up, and hands you the finished string. If a configured log file is missing, it self-heals to [None] and falls back to Manual rather than blocking the queue.

Inputs and outputs that matter

The one output you care about is final_prompt - wire it into Generation Core's positive_text. The rest of the fourteen outputs exist so Classic workflows can trace what actually ran: source_prompt, prompt_line, outfit_line, scene_line, the resolved indexes and counts (prompt_index_resolved, outfit_index_resolved, scene_index_resolved, prompt_count, outfit_count, scene_count), and the file stems (prompt_file, outfit_file, scene_file). Those feed Output Core's filenames so your saved files encode which outfit and scene made them.

Installing it

It's in the SickOllie pack:

cd ComfyUI/custom_nodes
git clone https://github.com/sickollie/SickOllie

or ComfyUI Manager → search "SickOllie". Restart and Ctrl+F5. Don't forget the starter logs: copy the SickOllieLogs folder from the pack's Starter Content into ComfyUI/input/ so the dropdowns have files to pick from.

Where people get burned

The classic trap is expecting a token to be replaced when it's not in the prompt. Prompt Core only substitutes what's there - if your manual prompt never mentions SCENE, the scene line won't insert. That's by design in Classic (Studio adds explicit placement modes); the fix is to put the tokens in your template. The other common hiccup is renaming a log file after selecting it - the selection goes stale, Prompt Core self-heals to [None], and you reselect. If a saved prompt looks wrong, check whether cleanup rules are eating something you actually wanted; that regex is powerful enough to mangle a prompt if you loosen it.