Cyberdelia Wildcard Prompts

The pack this node lives in is advertised as an LLM prompt engineer - it talks to LM Studio or Ollama and rewrites what you typed. Wildcard Prompts is the part that never calls anything. No server, no key, no model discovery. It reads templates you wrote, fills in your tags, and hands one over per run. It's the odd one out in the pack, and the one you'll actually leave in the graph.

Where it fits

Prompt variety has three established answers in ComfyUI: old A1111 bracket syntax ({brown|blonde|black}), which most people who arrived after 2024 never learned; __wildcard__ file processing, which a lot of workflows get from Impact Pack without realising it; and the current fashion of letting a local LLM rewrite your prompt until it stops being the same face.

Wildcards are the cheap one. Testing a character LoRA, a style LoRA or a fresh checkpoint means wanting the same trigger word against twelve scenes - cafe, rooftop, studio, golden hour - stored in a file you edit between runs instead of retyped in the browser.

How it works

Everything runs in order, inside one node. Your prompts text (or a .txt file) is split into blocks. {tag} placeholders are substituted from extra_tags, and a tag you never defined - or defined as empty - is deleted with the spacing cleaned up, so A portrait of a {type} woman. becomes A portrait of a woman., not A portrait of a woman. Then search_replace rules run in the order you wrote them, then prefix goes on the front.

The last step is what changes your workflow. mode: random prompt outputs a single-item list picked by random.Random(prompt_seed). mode: all prompts outputs every block as a list - and since prompts is the pack's only list output, one queue in that mode runs everything downstream once per prompt. Ten prompts, ten images, ten sampler passes; prompt_count tells you how many you just signed up for.

Selection uses its own random generator, seeded from prompt_seed alone, so rolling a prompt can't disturb your sampler's seed state. A fixed seed repeats the same pick, and the author's own note is that a random draw can land on a prompt you already saw.

The controls that matter

• prompts and extra_tags are the two required fields. Tags are one name=value per line (hair=blonde); lines starting with # are ignored.
• prompt_separator - blank line if a prompt can span several lines, one per line if each line is its own prompt.
• mode, prompt_seed - see above. Set prompt_seed's control_after_generate dropdown to randomize for a new draw each run.
• enabled is the switch you'll use most. Off, and the node sends fallback_prompt once, unchanged: templates, tags, search/replace, prefix and the file path are all skipped. Keep your normal prompt in there and you can A/B the wildcard against a fixed prompt without rewiring anything.

Outputs are prompts, preview (text only; in all-prompts mode, everything joined) and prompt_count. Wire prompts to the text input of Prompt Engineer - Text or Conditioning, or straight to an encoder. Keep Prompt Engineer in passthrough to test the selected prompt with no rewriting.

Install

Via ComfyUI Manager, search Cyberdelia, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/cyberdeliaAI/comfyui-cyberdelia-z-engineer.git
pip install -r comfyui-cyberdelia-z-engineer/requirements.txt

That requirements file is one line - requests - and it's for the pack's LLM nodes, not for this one. Coming from the standalone comfyui-wildcard-lora package: remove it, restart, hard-refresh the browser. The node ID was kept, so old workflows still load.

Where people get burned

It sends the same prompt every run. Classic ComfyUI caching - an unchanged node isn't re-executed, and this one reports prompt_seed as its change signal. Randomising the seed is what makes it re-roll; leave it fixed and you test exactly one prompt, forever.

prompt_file isn't the Load .txt button. The path is read on the ComfyUI machine, relative to ComfyUI/input, and must be .txt or it errors. The button reads a file in your browser, and it clears prompt_file when it does, folding the text into the workflow. Silent source swap.

Some mistakes stop the queue instead of warning. A tag name defined twice, a line missing its =, or a search/replace rule that empties a whole prompt all raise. Deliberate: a run that quietly produced eleven prompts instead of twelve is worse than one that stops.

Whole-word matching is on by default, so woman=male won't touch superwoman - I checked. Right default, and also why your replacement "doesn't work" when you wanted a substring; flip replace_match to literal text.

It doesn't expand {a|b|c}. Bracket syntax is a different mechanism, and these fields are flagged so the dynamic-prompts frontend extension leaves them alone. It also won't vary LoRA strength, steps or samplers - only text - so pairing a wildcard sweep with an X/Y plot is normal, not redundant.