Skip to content

Prompts for product agents

Copy a prompt and replace the brackets with product context.

01

Define the first job

A five-bullet pilot plan

Product: [product] User: [role] Job: [recurring task] Design one bounded pilot: trigger → tools → review → checked outcome. Return five bullets. Flag missing context instead of inventing details.

Read the guide
02

Choose the next tool

Three capabilities, backed by cases

Cases: [anonymized cases and their outcomes] Group misses by missing tool, wrong argument, stale context, or usability. Rank three capabilities by cases helped and review effort. Return one table with case IDs and evidence. Leave unsupported estimates blank.

Read the guide
03

Check the permission boundary

Concrete tests and expected results

Tool: [action and target objects] Actor: [acting identity] Review: [proposed confirmation] List authority, scope, and freshness checks at execution. Test wrong tenant, revoked access, stale arguments, duplicate approval, and a lost reply. Include the expected result and how to check each case.

Read the guide
04

Measure the pilot

Four metrics and a pilot decision

Job: [job] Baseline: [cases, verified outcomes, human minutes] Pilot: [same fields, plus review, rework, model and operating costs] Compare cost per verified outcome and released capacity. Separate observations from estimates. Return four metrics and one pilot decision.

Read the guide