Instructions

Instructions are an assistant’s baseline: how it behaves, what it’s for, and how it stays grounded. They load into every conversation, so they’re the right place for general behavior — not for the steps of one particular task.

Note

This field used to be called Prompt. It is called Instructions throughout Intric now — in the assistant wizard, in an assistant’s settings, and in widget settings. Nothing about what you write changes.

What instructions are for

Good instructions establish the assistant’s role, its tone, and the boundaries of what it should do. The clearer you are, the more consistent the results.

How to write them

Work through five questions when you draft instructions:

Task

What do you want the assistant to do? Describe it clearly and specifically.

Context

Why does this task matter? Give the assistant the background it needs to make good judgment calls.

Format

How should the answer look? Tell the assistant how to structure and format its response.

Persona

Who should the assistant act as? Define its role or area of expertise.

Tone

What tone should the answer have? Specify the style or feeling you want — formal, warm, concise.

Tip

Include negative instructions too — what the assistant must not do, such as “don’t make up facts you can’t find in the attached knowledge.”

Instructions or Skill?

Keep instructions general and put anything task-specific in a Skill instead. If you find yourself writing “when the user asks about X, do Y, then Z” in the Instructions field, that’s a Skill trying to get out. See Skills.

Iterating

Don’t aim to get it right on the first try. Write a draft, test the assistant, and revise. Ask the assistant to raise a follow-up question when something in your instructions is ambiguous — that’s often the fastest way to find the gap.

Common mistakes

• Writing task-specific steps into the instructions instead of creating a Skill for them.
• Being vague about tone or format, then correcting every answer by hand instead of fixing the instructions once.
• Forgetting negative instructions, so the assistant guesses at behavior you actually wanted to rule out.