croissant
Home / Blog / Planning

How to write a useful brief for an AI agent project

Croissant editorial team · · 2 min read

Give an implementation team the inputs, boundaries, examples, and outcomes needed to scope a practical agent workflow.

A good project brief describes the work, not a list of fashionable technologies. It should let someone unfamiliar with your business follow one request from arrival to completion. You do not need a technical specification to begin, but you do need concrete examples and a clear owner.

Describe one journey through the process

Start with the trigger and show the current steps. Include which tools people open, which information they copy, and where they decide what to do next. Explain the result the customer or team member expects, and where the process commonly gets stuck.

Provide examples and limits

Offer ordinary examples and the cases that require judgement. Remove sensitive information before sharing. State which actions the agent may take and which require approval. Include important exclusions, such as negotiating terms or changing records outside the agreed workflow.

A brief you can start with

Use this structure: “When this event arrives, our team currently follows these steps using these tools. We want this output. These examples show acceptable work. These cases need a person. This owner will review the pilot.” Add the information that would make the project unsuitable, such as unavailable integrations or a source that cannot be shared. A clear boundary helps an implementation team propose a realistic first stage instead of guessing what your broad objective means.

Define how you will accept the work

Agree a small test set and the evidence you need to see. Ask for a record of actions, a way to pause the workflow, and instructions for the person who will own it. Acceptance should cover recovery and handover, not only an attractive demonstration.

A reality check before you hit run

Anthropic recommends clear, well-documented interfaces for agent tools. Your brief should define what each tool is allowed to do. [1]

n8n shows how selected tool calls can pause for a person to approve or deny. Specify those decision points in the brief. [2]

Implementation checklist

  • Name the process owner and intended users.
  • List inputs, tools, outputs, and exceptions.
  • Include anonymised examples and approval boundaries.
  • Agree tests and the operational handover.

The sources behind the sauce.

Primary research and official documentation behind the cited points. The workflow examples and checklists are Croissant’s own guidance.

  1. Anthropic — Building effective agents
  2. n8n documentation — Human-in-the-loop for tools