What Are Instructions?
How system prompts shape AI behavior — identity, workflow logic, boundaries, and output format — and why instructions are the key to a genuinely useful agent.
Instructions are the missing piece between “AI that can access my tools” and “AI that actually helps me.”
You’ve learned that models process your messages and tools let AI take action. But without instructions, your agent has no idea what you actually want. It’s like hiring someone with all the right skills but never telling them the job.
How Instructions Work Under the Hood
When you chat with an agent, your message isn’t the only thing the model sees. Instructions get prepended to every single conversation. The model reads them first, before it ever sees what you typed.
This means instructions aren’t a one-time setup. They’re active context that shapes every response your agent gives. Change the instructions, and you fundamentally change how the agent behaves.
You’ll hear both terms used interchangeably. “System prompt” is the technical name. “Instructions” is what we call them in Gumloop because that’s what they are: you’re instructing your agent on how to behave.
The Four Things Instructions Control
Instructions can shape almost every aspect of agent behavior. Here’s how to think about what to include:
- Identity and persona: name, tone, communication style. Example: “You are Aria, a friendly customer support agent. Be warm but professional.”
- Workflow logic: steps to follow, order of operations, decision points. Example: “First check inventory levels. If stock is below 10 units, flag for review before confirming the order.”
- Boundaries and exceptions: what to refuse, edge cases, escalation triggers. Example: “Never share pricing without checking the latest price sheet. Escalate to a human if the customer mentions legal action.”
- Output format: how results are delivered and structured. Example: “Respond with bullet points. Keep responses under 100 words. End with a clear next step.”
Why Instructions Matter
Instructions are where your expertise lives. You know things about your business that no AI model was trained on. You know that “urgent” from Client A means “sometime this week” while “urgent” from Client B means “drop everything.”
Instructions let you transfer that knowledge to your agent. Without them, you’re asking a very capable system to guess at context it doesn’t have. Two teams can use the exact same model, connect the exact same tools, and get wildly different results. The difference is instructions.
You can’t instruct your way out of missing tools. If your agent needs CRM data and you haven’t connected your CRM, no instruction will fix that. Instructions guide behavior within the boundaries of what’s possible.
The Agent Formula
At this point, you have the complete picture:
| Component | Role |
|---|---|
| Model | Processes language, reasons through problems |
| Tools | Gives access to your systems and data |
| Instructions | Shapes behavior, encodes your process |
All three working together is what makes an agent actually useful.