Prompt Preset

If you've wired an LLM into ComfyUI, you've hit the same wall: the system prompt lives as a giant multiline blob inside one node, and the moment you want a second variant - "summarize" vs "tag extractor" vs "harsh critic" - you're either duplicating nodes or pasting text around like it's 2005. Prompt Preset fixes that by keeping prompts in an external .toml file with {{placeholders}} for the parts that change, then feeding the resolved prompts straight into the pack's chat nodes.

It's the companion to Chat Once and Vision Chat Once: those send, this one prepares. And unlike the chat nodes, this is the one place in the pack with a real dependency - it uses tomli to parse TOML.

How it works

Your file is a list of presets, each with a label, a system template, and a user template:

version = 1

[presets.summary]
label = "Summary"
system = """
You are a helpful assistant.
Style: {{style}}
Instruction:
{{instruction}}
"""
user = """
Summarize the following:

{{input}}
"""

The node reads preset_path, grabs the preset by preset_id, and does simple {{name}} string substitution. Two variable sources: input_text (becomes the input variable, and wins any name collision) and variables_toml - a flat TOML block for the rest:

style = "concise"
instruction = "Use short paragraphs."

It resolves both system and user templates, normalizes line endings to LF, and outputs the finished system_prompt and user_prompt - wire those straight into Chat Once.

Inputs that matter

• preset_path + preset_id - which file, which preset. Path must end in .toml.
• input_text - the changing content, mapped to {{input}}.
• variables_toml - flat key-value extras. Arrays, tables, and inline tables are errors here.
• keep_unresolved_variables (true) - leave unknown {{names}} in the output, or blank them out.
• fallback_user_prompt - used if preset_id doesn't match or the preset has no user field.

Outputs: system_prompt, user_prompt, plus preset_meta_json (metadata, including which variables were resolved) and preset_name (the label, or the id if there's no label).

Install

Part of ComfyUi_zaknak_nodes - ComfyUI Manager, or git clone https://github.com/zaknak/ComfyUi_zaknak_nodes into ComfyUI/custom_nodes, then restart. tomli is the pack's one declared dependency; Manager will pull it if it's missing.

Troubleshooting

• "toml parse error" / "unsupported extension": only .toml files, UTF-8, with version = 1 and a presets table at the root. JSON and YAML are right out - that's deliberate.
• "variables_toml" errors: it accepts only flat key-value pairs - strings, ints, floats, bools, and multiline strings. An array (tags = ["a","b"]) or a nested table is an explicit error, not a silent ignore.
• {{unknown}} survived into your prompt: that's keep_unresolved_variables working as designed. Unresolved variables are logged to the ComfyUI console (with which template they appeared in), so check there before blaming the model for rambling.
• Nothing resolves: preset_id must match a [presets.<id>] section exactly. If it doesn't and fallback_user_prompt is empty, the node fails loudly - which beats sending a half-rendered prompt to your LLM.