Your first AI team should be small.

Start with a finished thing

Choose a task whose result you can recognize: a brief based on three documents, a comparison of a few options, or a draft you can review. “Help with marketing” leaves too many decisions unstated. “Turn these interview notes into five themes, with a quote for each” gives the agent a clear finish line.

For a first test, use a copy of material you are comfortable sharing with your AI provider. Keep the task small enough to inspect the inputs and the result yourself. You are learning what the agent needs from you, not trying to automate an entire department.

Give the lead agent context

YankyBot starts with a Chief of Staff. Tell it what you are working toward, which files are relevant and what a useful output looks like. Add constraints that affect the work: your audience, the deadline, a length limit and any action that needs your approval.

Read these three interview notes. Group recurring problems into five themes, include the supporting quotes, and list anything the evidence does not resolve. Save a draft brief for me to review. Do not contact anyone.

This is an illustrative prompt, not a record of a completed task. Replace the inputs and boundaries with the ones that apply to your work.

Add a specialist when a job repeats

A separate agent becomes useful when it has a lasting responsibility. A writer can keep style guidance. A researcher can focus on evidence and source quality. Give each a specific role instead of creating several agents with the same broad instruction.

YankyBot agents can hand tasks to each other and return the result to the conversation. Start by asking the lead to delegate one bounded piece. More agents are not automatically more reliable: each handoff needs enough context and a result that the next person or agent can use.

Review the work, then improve the brief

Check whether the result answers your original question, whether its claims match the source material, and whether any step exceeded the scope. Correct the instruction that caused the problem. Save a stable preference in memory; keep a one-off correction in the task.

The next run should begin with a better brief, not a longer list of agents. Once a task works well by hand, you can consider a recurring schedule.