Give Your AI Team Clear Tasks and One Writer

Start with one visible deliverable

I used to hand over a whole requirement and find that fixing one part broke another. I now start by defining the deliverable. A first experiment can be one task with fixed inputs and a result you can inspect quickly.

For a comparison of three public product documents, I would ask for one table first. Research, proposal writing and sending a message each need their own decision point.

Fill five fields and name the boundaries

My five fields are function, input, output, boundaries and acceptance criteria. For a product comparison, I would replace “make a good report” with a specific extraction request that keeps missing information visible.

• Function: produce only the comparison table.
• Input: the three specified public documents.
• Output: product, feature, limitation and source columns.
• Boundaries: mark missing facts; do not guess or send anything.
• Acceptance: every factual row traces back to a source.

Choose roles before tools

A researcher returns sources and excerpts. A writer organizes confirmed material. A reviewer looks for errors and omissions. These are responsibilities; a small task can use one AI in separate stages.

When two agents work together, name one writer and give the other read-only review responsibility. Review comments should remain distinguishable from edits, so you know which version was checked.

Review against evidence

Ask a reviewer whether conclusions have sources, missing inputs could break the work, or a forbidden action occurred. Two models agreeing still leaves you responsible for checking the result.

Open the source material and inspect important fields. For code, inspect actual execution evidence. “Completed” is a claim; a deliverable with supporting checks gives you something to evaluate.

A reusable starting brief

“Read only the specified inputs. First list the task, inputs, output, files you must not touch and acceptance criteria. Ask about missing information. Once scope is confirmed, only you may edit the designated output; the other AI provides review comments. Return the result, sources and unresolved questions. Do not send anything externally.”

Run it once manually and refine ambiguous fields. The useful artifact is a work brief that still makes sense next time.