Lab 5: one brief, two AI vendors | The AI Agent Factory

In this lab, you write one brief for Brightline's payment-run proposal for Friday, October 23, and run it in Claude and in ChatGPT Work. You score both results, change the part of the brief responsible for the worst failure, if there is one, and split the work into two stages.

Where you work. Most of the lab is in a folder on your computer. Download brightline-lab-ch05.zip from the Labs companion, unzip it, and open the files in any text editor, such as Notepad or TextEdit. Only the runs use chats with Claude and ChatGPT, in steps 2 and 4. Keep your briefs and answers in the folder, not in a chat. They are yours, and Chapter 9 sets your saved brief to run on a schedule.

You can also do the lab with the Claude or ChatGPT desktop app. Open the folder in the app, and ask it to read LAB.md and start. The runs still happen in new chats.

How long. About 110 minutes in total. Each step ends with a file saved, so you can stop after any step.

The task. Six source files go with every run: the open invoice list, the vendor records, the approvals log, AP policy versions 3 and 2, and the text of one Tri-County Freight invoice. Five problems were put in them on purpose. They are the retired policy, an out-of-date payment terms line, a duplicate invoice, a Canadian-dollar invoice the policy does not cover, and a bank-change note inside an invoice. Dave, the controller, owns policy version 3 and approves the run. You write the brief and score the runs. The worker proposes, pays nothing and changes nothing.

What you do. Do the steps in order. This list says what each step is for. LAB.md, in the folder, gives the exact instructions, one Part for each step. Read each Part when you reach its step, not all at once. If a Part is unclear, upload LAB.md to a separate chat with Claude or ChatGPT. Ask it to explain the step you are on.

Predict (LAB.md Part A, 10 minutes, in the folder). Read Dave's one line and Maria's fourteen steps, in requests/. Write down which of the five problems each will miss, and why. You save: results/predictions.md.
Run (Part B, 35 minutes, in the folder, then in chats). Write your Four-Part Brief from the template in briefs/. Then Run A sends Dave's line, in a new chat on either AI vendor. Run B sends your Four-Part Brief in Claude. Run C sends the same brief in ChatGPT Work. For every run:
- Use a new conversation, and attach the same six files.
- Never attach the answer key.
- Switch off the memory feature first, so it does not change the test. In Claude, turn off Memory in the "+" menu. In ChatGPT, open a Temporary Chat and choose Unpersonalized.
- Record the model and its settings, such as the permission setting.
You save: briefs/payment-run-brief-v1.md, and every reply and file that comes back.
Investigate (Part C, 30 minutes, in the folder). Score each run with the rubric, the scoring sheet in answer-key/run-rubric.md. You may open it now, but not the answer key. Name the part of the brief responsible for each failure. Compare Runs B and C on facts, format, file destination and anything you had to adapt. Only then, open the answer key. You save: results/run-log.md and results/comparison.md.
Modify (Part D, 25 minutes, in the folder, then in a chat). Change the one part responsible for the worst failure and run it again. If your brief had no serious failure, record that instead of inventing one. Then split it into two stages with a stop after the exceptions list. You save: results/iteration-log.md and briefs/payment-run-brief-v2.md.
Make (Part E, 10 minutes, on paper or in a file). Write a Four-Part Brief for a task that repeats, in a role you know, and mark its controls. You save: your own brief, as briefs/my-own-brief.md.

One run shows how a worker behaved once. It does not prove that every difference came from the brief, and Dave's line may do better than you predicted. Run C needs a ChatGPT plan that includes Work. With free ChatGPT, use Chat and label it. If you use only one AI vendor, fill in the transfer plan, results/transfer-plan.md, instead. It asks what you would change in each part of the brief to run it on the other AI vendor.

Exam notes

These notes are not part of the lab. CCAO-F, Anthropic's Claude Certified Associate: Foundations, follows an exam guide written in July 2026, and its questions are about Claude only. Where this chapter teaches today's products and the guide expects a different answer, the note gives the answer that scores. Learn both. Certification and Portfolio Roadmaps has more.

Features. The exam guide expects you to pick a named feature: projects, research mode, chat or artifacts. It also names memory, skills, code execution and connectors, such as Google Drive and Gmail, and any of these can be the right answer. Today Claude often decides inside one conversation, with no mode to pick, and it adds tasks, Cowork and editors such as Claude Docs. On the exam, choose the named feature that fits the job. "Write a clearer brief" is not a feature choice. In the exam guide, "artifact" means Claude's Artifacts feature, not a Claude Doc or a Google file, even though Concept 5.7 uses the word for any file of its own. (D-003)
What a good prompt contains. The exam guide expects an effective prompt to give context, a clear task, the output format and constraints, and often a role, examples and a request to reason step by step. The Four-Part Brief covers the first four under its own names. On the exam, choose the option with the most useful context, a clear task, a format and examples. Asking Claude to work through a problem step by step is a valid technique, and "brief the outcome, not the steps" does not rule it out. (D-009)

Artifact checklist

results/predictions.md, with what you expected Dave's line and Maria's steps to miss
briefs/payment-run-brief-v1.md, your first Four-Part Brief, with each control marked
results/run-log.md, with Runs A, B and C scored and the part of the brief named for every failure
results/comparison.md, with Runs B and C compared on facts, format, file destination and changes needed, or results/transfer-plan.md if you have one AI vendor
results/iteration-log.md, with the one-part change, its rerun, and the two-stage version
briefs/payment-run-brief-v2.md, the saved brief that Chapter 9 will put on a schedule
A Four-Part Brief for one task that repeats, in a role you know, with its controls marked