🔱 Gemini Instruct
The free node that writes the other nodes' instructions
- system_instruction
Gemini Instruct is the quietest node in the Artha-Gemini pack, and also the one with the least excuse for drama: it doesn't call any API and needs no key. Its entire job is to build a system-instruction string that you feed into other Gemini nodes' system_instruction slots. Think of it as a reusable prompt template that lets you customize how every other node in the pack behaves.
That's more useful than it sounds. Most of the pack's nodes ship with a good built-in agent - Vision has its structured captioner, Cloth has its garment-lister, and so on. But the moment you want one of them to behave differently - "describe this image in the style of a film location scout" - you need a way to inject your own instructions. That's the slot this node feeds.
How it works
Three multiline text inputs, no logic beyond string assembly:
- role - who the agent is. Defaults to
You are an intelligent ai assistant. - task - what it should do. Defaults to
Your task is to... - instructions - any extra guidance.
The node strips double quotes from all three fields (a small safety touch), then assembles them into the pack's system-instruction format: Role: ... on its own line, then Task: ..., then the instructions. The output type is ARTHAINSTRUCT, the custom type that the pack's API nodes accept in their system_instruction inputs - so it's designed to plug straight in.
No API block here, no model, no temperature. You're just shaping a string.
Inputs and outputs
- role, task, instructions - the three text fields above.
- Output:
system_instruction(ARTHAINSTRUCT).
Connect the output to the system_instruction input on Gemini Vision, Motion, Prompter, Imagen, or any other pack node that exposes that optional slot. If a node has that input and you don't wire it, it falls back to its bundled agent - so Instruct is purely additive: wire it in when you want control, leave it out when the default is fine.
Installing it
Via ComfyUI Manager (search Artha-Gemini) or:
cd ComfyUI/custom_nodes
git clone https://github.com/Cyrostar/ComfyUI-Artha-Gemini
cd ComfyUI-Artha-Gemini
pip install -r requirements.txt
No model downloads, no API key, no cost - this node never leaves your machine.
Where people get burned
The main failure mode is expecting it to do more than it does. Instruct doesn't prompt Gemini itself; it just produces the instruction text. If you wire Instruct straight to a KSampler or expect it to generate anything, nothing happens. It's a text producer for other nodes - the dependency flows through the pack's system_instruction inputs.
Second, remember it overrides the built-in agent. If you wire in a vague role and task, you're replacing a tuned, purpose-built instruction with your own, often worse, one. A good pattern is to start without Instruct, see how the default agent performs, and only add custom instructions when you know what you want changed. And a small practical note: the ARTHAINSTRUCT output only matches ARTHAINSTRUCT-typed inputs, so it won't plug into arbitrary STRING slots elsewhere in ComfyUI.
Inputs (3)
| Name | Type | Default | Description |
|---|---|---|---|
| role | STRING | You are an intelligent ai assistant. | — |
| task | STRING | Your task is to... | — |
| instructions | STRING | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| system_instruction | ARTHAINSTRUCT | — |