Appendix: practical run sheet | The AI Agent Factory

This appendix is for the editors, and for an instructor who wants to test the case on the product. The practical has not yet been run. Until it is, the run times, the line counts and the run notes in the exhibits are illustrative.

What the practical is for

The case joins parts that were each checked against Anthropic's help pages. It has not been checked that they join in the way the story says. This sheet tests the joins.

Setup

A Claude Team plan with three members, standing for Nisha, Sana and Hamza.
A shared Google Drive folder, Lakeshore/2027-09, connected to Claude.
The Claude desktop app on one computer, with a local folder Lakeshore allowed.
A test Slack workspace with Claude Tag, which is Claude in Slack, in one channel, as in Case 3's run sheet.
A test mailbox outside the organization, standing for Dana.

The data

Make two files of about 3,400 lines each, with the same totals as the case: a bank file that closes at $201,480.00 and a ledger that closes at $198,300.00, with two unmatched deposits of $6,120.00 in all and seven unmatched checks of $9,300.00 in all. Make a second bank file with one more line: a returned check, minus $3,450.00, dated September 30. The lab's small files have the same totals and can be used for a first pass.

The tests

Test	Do this	Record
1	Make the project, with the instructions in Exhibit 3 and the skill in Exhibit 4	Screens of both
2	In a chat in the project, ask Claude to list the files in the drive folder	Whether a project chat can read the connected drive
3	Send the request in Exhibit 5	The time. Whether Claude ran a program. The number of pairs. The figure.
4	Look at the four pieces	Whether the run updated the linked pieces or made new ones
5	Look for the run note and the matched list in the drive	Whether Claude could save them there
6	Add the sentence about slide 2 and run again	The time. Whether only slide 2 changed.
7	Give the worker in Slack the morning check as a scheduled task. Put the second bank file in place.	What the worker posts, and whether it changed anything
8	Send the second request	The time. The memo's first line. Whether all pieces agree.
9	Export the cover, the memo and the slides as PDFs and email them to the test mailbox	That the files open, and what they show
10	Share the project, the skill and the four pieces with the two other members. As one of them, send the request.	Whether a second member's run uses the same rules, skill and pieces
11	Ask the worker to read the latest run note in the drive and compare it with a fresh download	Whether the worker can read it
12	Make a sheet from the month's 6,800 lines	Whether the sheet holds them, and how long it takes
13	As Owner, turn on only the Drive connector. As a member, try to send an email from a chat in the project.	That the chat cannot send
14	Do the lab's core path on a free account	Each step that does not work as written, and the minutes taken
15	In a desktop chat, read a file from the local folder. Then try to reach the same folder from a scheduled cloud task.	Whether each one can, and whether the cloud task still can with the app closed

What to measure

Minutes for each run, and for each person's checks.
For test 3, check twenty pairs by hand, and count any wrong ones.
For tests 4 and 8, the figure in each piece.

What to keep

The project's instructions and the skill, as text
Each run's chat, exported
The two run notes
Screens of the four pieces after tests 3, 6 and 8
The worker's message from test 7
The PDFs from test 9

After the practical

Replace each illustrative label in the exhibits with what was measured, or correct the story. If a join does not work as the story says, change the story to what the product does, and say in the exhibits what was changed.