Multimodal — System prompter

The main LLM node has a system_prompt input, and it's just an empty multiline box. Type your own prompt engineering essay into it, every time. This node is the pack's answer to that: a curated dropdown of ready-made system prompts - image describer, SDXL prompt writer, Flux prompt writer, QC reviewer, edit director, context updater, strict JSON schema - plus the ability to load your own from a file or paste custom text. The output feeds straight into the LLM node's system_prompt, and it's the difference between a blank box and a coherent persona in one click.

The inputs that matter

selection is the top switch with three modes:

• preset - the built-in library, chosen via preset_file. The defaults (01_image_describer, 02_prompt_writer_sdxl, 03_prompt_writer_flux, and so on through the pack's presets/system/*.md files) cover the most common jobs, and they're real Markdown files on disk, not strings baked into code.
• from_file - load any .md/.txt path via file_path.
• custom - paste your own prompt into custom.

preview_chars (default 1200) controls how much of the prompt the system_prompt_preview output shows; set 0 for the whole thing. custom_empty_when_not_selected is a small UI courtesy - it keeps the custom field visually clean unless you're actually in custom mode.

Two outputs: system_prompt (wire to the LLM node) and system_prompt_preview (wire to a Show Text node so you can read what's about to be sent without expanding the node). It's a boring but very practical division: the value travels, and the preview stays for eyeballs.

The preset trick you'll actually use

The preset dropdown isn't static. It scans presets/system/ at load time, so you can drop your own .md files in there and they appear as new options - a personal system-prompt library that lives with the pack and survives restarts (a restart is required to refresh the list). The included presets are genuinely good starting points: the prompt writers know the difference between SDXL and Flux phrasing, and the QC reviewer is built to critique an image against a stated goal and emit edit instructions, which slots straight into the closed-loop edit workflows this pack leans on.

One thing to keep straight: this node builds the system prompt, it doesn't consume one. It has no idea what images you've loaded, and it won't update your context - that's the Context handler's job. For the common "give the model a professional identity" task, this is the one you reach for. It's also the least fiddly node in the pack, which after the wheel saga is a relief.