Chapter 1 · Lesson 3
Request an Output You Can Use
Design the answer’s structure so it can move directly into the next step.
AI gives you a thoughtful wall of text, but you still spend ten minutes turning it into a plan.
Format is part of the task
Do not treat structure as cosmetic. A table supports comparison, a checklist supports execution, and labelled fields can feed another tool. Ask for the shape that matches what happens next rather than accepting a generic essay.
Define fields and their purpose
For repeatable work, name the fields you need and what belongs in each one. A risk register might require Risk, Evidence, Likelihood, Impact, Owner, and Next action. Clear fields make missing information visible instead of hiding it inside fluent prose.
Use machine-readable output carefully
JSON or CSV can help when an answer will enter software, but valid syntax does not guarantee valid facts. Validate required fields, allowed values, dates, and source fidelity before an automated step uses the result.
Build the mental model
The best output format is the one that reduces work and risk in the next step.
Design backwards from use
Start with what happens after the answer. A decision needs options and trade-offs; execution needs owner and due date; automation needs validated fields and allowed values.
Represent missingness explicitly
Define a value such as “not stated” or null. Otherwise a fluent model may silently complete missing fields, making invented information look operationally ready.
Meeting notes must become tasks in a project tracker.
Summarise these meeting notes and list the actions.
Return one row per action with: Action, Owner, Due date, Evidence quote, Confidence (high/medium/low), and Open question. Use “not stated” for missing owner or date. Do not infer commitments. Sort low-confidence rows first for review.
The answer now maps to the tracker, preserves evidence, exposes uncertainty, and puts the highest-review items first.
When to use prose, tables, or structured data
Use prose for explanation and narrative, a table for comparison and review, a checklist for execution, and JSON or CSV only when software will consume the result.
Before automating structured output, validate required fields, data types, allowed values, and relationships. Correct syntax is not the same as correct content.
Specify an output for the next tool or person
Pick a recurring answer you currently reformat. Define its fields, missing-value rule, and review order.
Return [format] with fields: [...]. For missing information use [...]. Validate [...]. Sort or group by [...].prompt choice
Which output request is most useful for turning meeting notes into action?
Compare the prompts closely. Look for useful context, constraints, and a clear output.
Ask for the structure required by the next step, not merely a shorter answer.