Best practices: writing effective prompts and a testing checklist

How to write your AI agent's system instructions and a checklist to test them before publishing.

Jun 26, 2026

Watch this step in the video · How to write good instructions

The System instructions (the prompt) are what define how your AI agent thinks, replies and acts. A good prompt is the difference between a bot that confuses customers and one that solves their problem. This guide gathers concrete practices for writing effective prompts in Omnifox, plus a checklist to test them before publishing.

Where the prompt lives

Open your agent under AI Agents and go to the Instructions tab. There you'll find the System instructions field, with instruction Templates (Sales, Support, FAQ, Reception and more), insertable Variables and the Optimize with AI button.

How to write a good prompt

  1. State the role and goal in the first line. E.g.: "You are the sales assistant for [company]. Your goal is to qualify the lead and book a demo."
  2. Give business context: what the company sells, hours, policies, what it must NOT promise.
  3. Set clear boundaries. Tell it what to do when it doesn't know something: "If you don't have the answer, offer to hand off to a human" instead of making things up.
  4. Use variables instead of hard-coded text: insert Contact name, Language, Channel, Today's date from the Variables panel. The message personalizes itself.
  5. Specify tone and format: short replies, no jargon, one question at a time. Tone is also adjustable on the Behavior tab.
  6. List the available actions. If the agent has tools (book appointment, update contact, transfer), explain when to use each.
  7. Avoid contradictory instructions. "Be brief" and "explain in detail" in the same prompt confuse the model.

Once you have a draft, use Optimize with AI: it generates an optimized version of the prompt that you can review and apply.

Testing checklist (before publishing)

Use the agent's Sandbox (Live tester). Walk through this list:

  • Greeting and happy path: the agent introduces itself well and solves the typical request.
  • Out-of-scope question: does it hand off to a human instead of inventing an answer?
  • Missing data: does it ask for the info it needs before acting?
  • Personalization: load a real contact with Load real contact and confirm it uses their name/language.
  • Tools: does it fire the right action (book, tag, transfer) at the right moment?
  • Tone and language: does it answer in the customer's language and with the defined tone?
  • Hostile or ambiguous messages: does it stay calm and in role?
  • Reply length: does it respect the Max tokens per reply limit?

Use Reset conversation between scenarios so leftover history doesn't pollute the test.

If it doesn't work

  • The agent ignores an instruction: move it higher and make it more explicit; models prioritize what comes first.
  • It hallucinates data: add an anti-hallucination rule and connect a Knowledge base.
  • Replies are too long: lower Max tokens per reply on the Behavior tab.

Tips

  • Change one thing at a time and re-test: that way you know what caused the improvement.
  • Save versions: the History on the Instructions tab lets you compare and roll back.

Related articles

Was this helpful?

Related articles

Message us on WhatsApp