Imperium PostBooks for independent minds

Free companion worksheet

One task.
One clear test.

Use this worksheet to plan a small AI trial you can actually judge. Begin with work you understand well enough to check.

Original companion resource · Imperium Post · 2026-09-25

Skip to the blank worksheet

The print button needs JavaScript. You can also use your browser’s Print command. Nothing you write on a printed copy is collected by this page.

A worked example: a fictional review reply

Imagine a café practising replies to routine positive feedback. This example uses a made-up business and review. It is a practice exercise, not a reported customer result or an output claimed to have been tested with a particular AI model.

1. Start with the task, not the tool.

Task: draft a short, warm response to a positive review. Keep human: verifying what happened, deciding whether the reply fits, and publishing it. A complaint, refund demand or dispute is a different task and stays with a responsible person.

2. Give the draft a boundary.

A vague request such as “write a reply to this customer” leaves the tone, length and promises open. A more specific practice instruction is:

Draft a reply to this fictional positive review for a fictional café.
Review: “Friendly service and a lovely cup of tea. I’ll come back.”
Tone: warm, plain and brief. Maximum 35 words.
Thank the reviewer. Do not add discounts, promises, personal details,
product claims or facts that are not in the review.
Return only a draft for a person to check. Do not publish anything.

3. Know what an acceptable draft looks like.

Thank you for the kind words. We’re glad you enjoyed the tea and the service, and we’d be happy to welcome you back.

Illustrative draft: written for this worksheet, not evidence of a model’s reliability. The draft stays within the supplied facts. It offers no reward, uses no private details and makes no unsupported promise.

4. Check the failure case.

Suppose a draft adds “Come back tomorrow for a free pastry.” The tone may sound friendly, but the offer was never authorized. Delete it. A useful checking routine looks for added commitments as well as spelling mistakes.

5. Measure the whole job.

Record your ordinary drafting time. Then record preparation, prompting, checking and corrections in the trial. If review makes the job slower or less reliable, change the process or stop. A fluent answer is not the success measure.

Before real use: do not paste real customer or employee information, private business data or confidential material into a tool just because the practice worked. Establish an appropriate data process first. Keep the fictional reply as practice; do not post it as a response to a real customer.

Your trial card

Use one card per task. Write the stop rule before the first trial.

The recurring task

What specific draft or transformation needs doing? How often?

The current process

Who does it now? What does a satisfactory result look like? Record actual time if known.

The input boundary

What may go into the tool? Use fictional practice data first. What must stay out?

The output instruction

State the task, facts, tone, format and things it must not invent.

The human check

Who checks every output? Against which original source or policy?

The cost of a mistake

Could you detect the error? What would happen if you missed it?

The measurement

Record preparation + tool use + review + corrections + recurring fees.

The decision date and stop rule

When will you review the trial? What would make you stop immediately?

The verdict

Keep, change or stop? Base the decision on the completed work, errors and total effort—not how impressive the draft sounded.

Count the cost properly.

Before adding a subscription, use our guide to the full cost of an AI tool. It separates subscription fees from setup, review and rework.

Make your next task the starting point.

Explore the complete 14-chapter guide to customer work, marketing, admin and the checks that keep you in control.