Write instructions for your Rovo agent

Instructions are used when creating (or editing) an agent to customize it to your needs. They define things like:

• The expectations or the purpose of your agent.

• The limitations - what you’d like the agent to do and not do.

• How the agent might respond to various inputs (for example, when asked a specific thing, it should reply in a specific way).

• How the agent should interact with people (for example, you may want your agent to have a particular tone - like “always respond like a pirate”).

If you’re looking for tips on chatting with agents and the kinds of prompts you’d write day-to-day, see Chat with an agent.

You provide instructions in a few key places in your agent.

• To your agent, in the Instructions field, to provide it with goals, behaviors, procedures, etc.

• Within the Trigger and Instructions fields of a Subagent, so your agent knows when to hand off a task to a subagent, and the subagent knows what to do.

Tips and best practices

Writing good instructions is an art, and you’ll most likely need to iterate on your instructions a few times to get your agent working just right.

To make your instructions more effective:

1. Keep your instructions relatively short to start with

   1. Shorter prompts are easier to iterate on (you can track how small changes to the prompt improve or worsen the agent’s performance). It can be harder to troubleshoot or iterate on a longer prompt.

   2. Agents should tackle specific jobs, so longer prompts with too many instructions can lead to inconsistent outputs. This is because, rather than actioning the full list of instructions, the agent will choose parts to prioritize.

2. Provide a role, tasks, and the relevant context to completing that job

   1. Role: The agent’s role is the name for the job you’re giving them. For example, if you’re making an agent to help you make decisions on projects, you might say:
      “You are a project manager who’s great at making unbiased decisions.”

   2. Job: Jobs can be written as one or more ways you expect your agent to help people. Jobs could be divided up amongst subagents or contained within the agent's instructions. To use the previous example, you might say:
      “You have the following jobs: Reviewing an existing decision, Finding related decisions, Providing additional resources or best practices to help with effective decision making.”

   3. Context: This is where you can choose to go very detailed, or keep it light. Context is essentially any extra detail you think your agent might need to deliver on its jobs. Context could include referencing the agent's knowledge or tools, and giving specific examples with example outputs.