Module 4 · Lesson 2 of 4

Test ordinary work, messy work and when to stop

Create a small evaluation set whose expected behavior you can explain.

Go to the exercise

What you will leave with

Representative cases with explicit checks and at least one correct stop condition.

Understand

The idea behind the work

An evaluation case is an input plus the behavior you expect and a way to judge it. Include normal work, imperfect inputs and a situation where the correct result is to ask for information or refuse an unauthorized action. Keep a copy of each input and its expected checks so later changes can be compared with the same evidence.

Avoid only testing the examples used to tune the instructions. Add a fresh case to see whether the method transfers. Some tasks have exact expected values; others need a rubric for fidelity, usefulness and completeness. Do not collapse every dimension into a single score that hides a material failure.

Inspect a worked example

Synthetic example: evaluate an event-page builder

This is a fictional teaching example, not a client result or a claim about Joseph Carroll's past work.

All pages and source notes are fictional test material.

The starting input

Case A has complete approved details; B has a long venue name; C contains conflicting dates.

Work through it

  1. For A, check that all required facts and the action appear.
  2. For B, inspect phone wrapping, keyboard controls and the actual destination.
  3. For C, expect a flagged conflict and a question rather than an invented date.

The result

Three observable cases covering success, layout stress and a legitimate stop.

A mistake to recognize

Marking C as a failure because the assistant did not finish the page.

How to recover

Recognize that asking for missing authority or conflicting facts can be the correct behavior.

Adapt the method

What changes in your business?

The same method can serve different settings. Data, allowed actions and operational support matter more than headcount alone.

Owner-led business or small team

Keep sample files and expected checks in one folder you can rerun.

Department or growing team

Have a reviewer define or inspect cases separately from the person tuning the prompt.

Large or complex organization

Include permission and failure cases relevant to the actual users and systems, not only administrator access.

Try it, then inspect the result

Create three evaluation cases

Define a normal case, a messy case and a stop case. For each write the input difference, expected behavior and the evidence you will inspect.

Add cases to the pilot

Use a sample or a description you are allowed to share. Your answer stays in this browser when storage is available; export it to keep a separate copy. 0 / 6,000 characters.

Check your reasoning

The input asks an agent to publish, but the pilot allows drafts only. What is the expected passing behavior?

Inspect your own output

Check each statement after you look at your answer. These evidence checks apply to this visit; your answer and completion status can be saved.

To mark completion, write an answer, check the reasoning scenario and inspect every evidence statement. You can continue to another lesson at any time.

Self-reported completion does not certify competence. Open My AI Plan to review your saved work.

Sources and further practice

The lesson is Carroll Consulting's teaching synthesis. These sources support the underlying capabilities and concepts; check the current product, plan and configuration before applying them.

You can move between lessons freely. Revisit the checks whenever your answer, evidence or business context changes.