How to write instructions for an AI agent
Writing instructions for an AI agent is the single factor that most determines whether it is useful or embarrassing. In Jestor, an agent carries its own instructions, a knowledge base and a set of fields it may edit, and the instructions are what turn a general model into something that behaves like your company.
Why vague instructions produce vague agents
"Answer customer questions politely" gives the agent no boundary. It will answer questions it should escalate, invent details it does not have, and adopt a tone that is not yours.
The fix is not a longer prompt. It is a specific one, with explicit limits.
What good instructions define:
- The exact scope of what the agent handles
- What it must refuse or escalate
- The tone and level of formality
- What to do when the knowledge base does not cover something
- Which information it must never state without a source
Structure of a reliable instruction set
| Element | What it establishes | Example of a weak version |
|---|---|---|
| Role | What the agent is for | You are a helpful assistant |
| Scope | What it handles | Answer anything |
| Refusals | What it must not do | Unstated |
| Escalation | When a human takes over | Unstated |
| Uncertainty rule | What to say when unsure | Unstated, so it guesses |
| Tone | How it sounds | Be professional |
Step by step to write them
- State the role narrowly, naming the process rather than the company.
- List what is in scope, explicitly.
- List what is out of scope, and what to do instead.
- Write the uncertainty rule, such as saying it does not have that information and escalating.
- Attach the knowledge base, and tell the agent to answer from it.
- Select only the fields it may edit, keeping the scope narrow at first.
- Test with the awkward cases, not the easy ones.
Why the uncertainty rule matters most
An agent without an explicit instruction about uncertainty will fill the gap, because producing an answer is what it does. Telling it exactly what to say when the knowledge base does not cover something converts the most damaging failure mode into a routine escalation.
Why choose Jestor
In Jestor, an agent carries instructions, a knowledge base from your documentation or public links, and an explicit list of the fields it can change. It can read attachments with OCR, classify, move cards and close them within those limits, with human in the loop available and a record of what it told each customer.
Frequently Asked Questions
What is the most important instruction to include?
What the agent should do when it does not know, since that is where the worst failures happen. See jestor.com.
Should one agent handle everything?
No. Narrow agents per process are more reliable than one general one.
Can I limit what the agent changes?
Yes. You select exactly which fields it has permission to edit.
Video Tutorial: Step by Step
With Jestor, you can automate workflows, connect teams, and build internal systems your way, all without code and powered by AI. Discover Jestor at jestor.com and see how to take your company's operations to a new level of efficiency and control.