Give the agent a bounded task

What you'll do

Write a bounded prompt for an agent-assisted BI task and assess a synthetic response. Specify evidence, allowed effects and a stopping point. Your output is a prompt/review transcript using synthetic metadata; no client data or real service credentials are required.

Ask for an artifact with a boundary

A useful task names the decision, the inputs and the result a reviewer can inspect. "Build the analytics" leaves grain, source authority, deployment and acceptance implicit. "Draft a Sales grain brief from these approved metadata files" identifies a local artifact and preserves review before downstream effect.

Separate what the agent can infer from what a person must decide. The assistant may identify a candidate key or explain a possible join loss. It cannot invent an authorised grain confirmation, independent business expectation or target permission. Ask it to surface uncertainty rather than fill gaps with plausible values.

Write the task in plain language

The following is a prompt example, not a shell command or permission to access a real source:

Using only the approved synthetic index and draft Sales seed,
draft a fact-grain brief for review. Identify candidate keys,
join risks and unresolved population rules with evidence references.
Do not invent expected values or human approval.
Stop with the draft and the specific decisions a reviewer must make.

Ask for references to the actual inputs. A convincing paragraph without evidence can hide an unsupported assumption. The reference need not expose a private file to the academy; the local reviewer should be able to trace the claim within approved custody.

Test the response against a failure case

Suppose the agent drafts one fact row per order line but selects a header total as the amount. Ask it to explain what happens when one order has multiple lines. The correct next step is to expose the grain mismatch and request the owning design decision. It should not quietly invent an allocation rule that makes one synthetic total match.

Likewise, a missing baseline in a report task should remain unavailable. A measure name alone cannot authorise positive sentiment. A generated semantic model does not imply refresh or correct deployed results. A good prompt makes these boundaries easier to review, but the actual response still needs inspection.

Keep scope through the next step

If the draft is useful, name the next separately authorised action. Validating a draft, ratifying grain, reading a source and deploying a definition are different operations. Do not let a general request to help with a model silently expand into all four.

Record both the request and the reviewed response in your synthetic transcript. Annotate one supported inference, one unresolved business decision and one proposed action that would require further scope or consent. Include the result's limitations in the final handoff rather than leaving them buried in an intermediate message.

Your expected output is a bounded prompt and an evidence-based critique. This exercise does not require a live agent or establish that any model is qualified. If you use an external provider, synthetic inputs still make the boundary explicit; client evidence requires its own approved handling. Learning to frame the task is useful, but it does not transfer business acceptance to the assistant.

Check your understanding

1. Which words in your prompt prevent an unreviewed source or deployment action?
2. How would you detect an unsupported allocation rule in the response?
3. What should the agent return when a required business definition is missing?