OpenAI System Instruction

A system instruction is the closest thing a GPT model has to a job description: it's the standing rules that apply before you ask anything. OpenAI System Instruction is a tiny pass-through node that attaches one of those rule sets to a client, so every downstream OpenAI node in the workflow behaves according to the same brief - no copying the instruction into every prompt box.

It's part of the ERPK Collection, the community pack wrapping the OpenAI API for ComfyUI. This is a configuration utility, not a generation node: its whole job is to dress up a client and hand it along.

How it works

It takes a client (from OpenAI API Config) and a system_instruction string, and it outputs the same client with the instruction baked in. You chain it:

OpenAI API Config → OpenAI System Instruction → OpenAI Vision / Text / Image nodes

The wiring order matters: anything downstream of this node inherits the instruction; anything upstream ignores it. That's how you build a workflow where the image critic "always answers in one sentence" while the text generator right next to it has a totally different brief - just put them on different sides of the node.

The system_instruction input is required and multiline, so you can paste in a real prompt-engineered paragraph: role ("you are a senior art director"), rules ("never mention the watermark"), output format ("respond as JSON with keys...").

The inputs that matter

• client - the OpenAI client to configure. Required. Comes from OpenAI API Config.
• system_instruction - the standing behavior. Required. This is where the actual work of your workflow lives: write it well and every downstream node benefits.

Output is a single client socket - same type, now configured.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk
pip install -r requirements.txt

Or ComfyUI Manager → search erpk → ERPK Custom Nodes. No key setup specific to this node; the client it receives already carries your credentials from OpenAI API Config / ERPK Settings.

Common issues

• Instruction seemingly ignored - check the chain order. If a Vision node is wired directly to API Config and bypasses this node, it never sees the instruction.
• Conflicting instructions - a system instruction sets the baseline, but per-node prompts still win day-to-day. If a downstream prompt contradicts it, the prompt usually takes precedence in practice.
• It's per-client, not global - re-using the same Config node elsewhere (for a different branch) won't inherit this node's instruction unless you route through it. That's a feature: it keeps different branches isolated.

Honestly, this node earns its keep the moment your workflow grows past one OpenAI node. Two branches, two personalities, one Config node upstream - that's the scenario it exists for. If you're running a single text node, skip it; you can paste the instruction into the prompt.