Prompt Rule Composer

Here's the automation pattern that keeps coming up: you have a caption - from a dataset, an image tagger, or an upstream model - and you want to turn it into a prompt by applying a set of human-readable rules. Prompt Rule Composer is a tiny rule engine for exactly that. Write rules like girl(!dress) -> casual girl or /red|blue/ -> colorful, feed it a caption, and it appends the rule's output whenever the left side matches. It's the pack's most feature-rich node, and honestly the most interesting one.

It's from artyclaw/artyclaw-comfy. If you're building any kind of batch caption-to-prompt pipeline - enrichment, style transfer, dataset tagging - this is the node you'll actually reach for.

How it works

Rules live either in the rules_text box (multiline) or in an external rules_file, and the syntax is three flavors:

• Literal terms: girl, full body -> feminine - matches if the caption contains girl and full body.
• Slash-regex: /red|blue/ -> colorful - matches if the caption matches the regex.
• Negative-lookahead: girl(!dress,!skirt) -> casual outfit - matches girl only when the excluded terms are absent.

When a rule's left side matches, the right side is appended to the output. A few behaviors matter: dedupe_appends stops the same phrase being added twice; consume_terms removes matched terms from the working caption (so later rules don't re-fire on them); consume_all_occurrences removes every instance rather than one; word_boundary forces whole-word literal matches; case_insensitive is on by default; and max_appends caps the number of additions. All of these feed the standard flags Python's re module provides, and the rule parser handles comments (#) and blank lines.

Inputs and outputs

• caption - the input text (required).
• rules_file / rules_text - where the rules come from; the file wins if both are set.
• case_insensitive, word_boundary, dedupe_appends, max_appends, joiner, output_only_new, consume_terms, consume_all_occurrences - the tuning knobs listed above.

Three outputs: prompt (the composed result), matched_rules (which right-hand phrases fired), and matched_rules_full (the raw matched rule lines) - handy for debugging exactly why a caption got what it got.

Install

ComfyUI Manager → search ArtyClaw Comfy Nodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/artyclaw/artyclaw-comfy

Restart. Standard library (os, re) only.

Where people get burned

Order matters. Rules are applied top-to-bottom, and because consume_terms removes matched text as you go, a rule higher in the file can strip a term that a lower rule wanted. If rules seem to "not fire," check whether an earlier rule consumed the term. Second, the max_appends default is 0, which means unlimited - most people expect 0 to mean "none," but here it's off. Third, output_only_new changes the output from caption+appends to just the appends; flip it by accident and your base caption vanishes. And the rules-file path resolves against the working directory, same as Prompt Block Loader. It's the node with the most to learn, but the rule syntax is genuinely readable - which is the point.