# Teaching note (for instructors) (/claude-and-chatgpt-2/case-4-nishas-month-end/teaching-note-for-instructors)

---
type: Document
title: "Teaching note (for instructors)"
description: "For instructors: learning objectives, a 75-minute discussion plan built on the six core questions, a board plan with a decision layout, notes on each question, takeaways, two objections students raise, what is assumed and not observed, lab grading, and the final assessment."
status: stable
order: 0.87
ksor:
  owner: team:panaversity
  audience: [ public ]
  approval:
    by: process:panaversity
    at: 2026-11-12T13:13:29Z
chapter: "F00G"
part: Front
expert_status: required
concepts: []
last_verified: 2026-11-12
sources:
  - id: run-sheet
    title: "Nisha's Month-End: practical run sheet (the practical has not yet been run)"
    resource: "appendix-practical-run-sheet.md"
generated:
  at: 2026-11-12T13:13:29Z
  by: process:panaversity
trust_tier: unverified
build_id: sha256:ee289d75a60f3acc7dc4cab134056c3f9c1ab9c4abdba810061b7f3810eb4d15
dirty: true
ksor_version: 0.0.60
---

This note is for instructors. Students should read the case and the exhibits first.

## What the case is about

A firm of three has every part of the Claude app and uses each one alone. A monthly job forces them to join the parts: a project, a skill, a connected drive, a folder, the temporary computer in the cloud, artifacts and a scheduled task. Two days of work become two hours. When the bank's figures change, the setup notices first and does the job again.

The case is the capstone of this part of the book. Its claim is that the parts are more useful joined than alone, and that joining them moves people's time from doing the work to answering for it. The closing decision asks how far that may go.

## Learning objectives

After this case, a student can:

1. choose the right part for each need, and say why;
2. join the parts into one setup for a job that repeats;
3. say what the setup does and what stays with a person;
4. fix the setup, not the result, when something is wrong;
5. place files where the job that needs them can reach them;
6. weigh how far a setup may run alone, and what the work is worth when preparing it becomes cheap;
7. answer the four questions of the series for a job of their own: is the work good, whose name is on it, whose word does the worker take, and what stays with people.

## Before class

- Students read the six story pages and the exhibits, about 50 minutes. The lab takes 75 to 110 minutes more.
- The case stands alone. The overview's part *You do not need the earlier cases* gives a new reader what they need.
- If students do the lab, they should do it after part 3 and before part 4.

## Discussion plan (75 minutes)

| Minutes | Question | Aim |
|---|---|---|
| 0 to 10 | Core 1. Where did the 17 hours go? | The time was in matching and copying, not in writing. A better editor would not have saved it. |
| 10 to 25 | Core 2. Take one part away. | Each part does one thing. Wednesday needs the schedule, the drive, the skill's rule and the links. |
| 25 to 40 | Core 3. Fix the setup, not the result. | A fix by hand reaches one piece for one month. Students should also find when a fix by hand is right. |
| 40 to 50 | Core 4 and 5. What could each approver know, and which test was wrong? | Approval rests on tests done without the AI. The setup makes them faster, not optional. |
| 50 to 70 | Core 6. What does Nisha tell Walter? | A vote on the line, then on the price. Then: what does your answer still trust? |
| 70 to 75 | Wrap-up | The four rules |

## Board plan

Draw two columns early and keep them up: **the setup did** and **people did**. Fill them from the students' answers to core 1 and 2. By core 4 the right column should hold only checking, approving, deciding and telling.

### Decision layout for core 6

| | Prepare only | Send if READY | Send always |
|---|---|---|---|
| Charge for the hours | | | |
| Charge for the pack | | | |
| Charge for the service | | | |

Ask each student to put a mark in one square. Then ask the students in the most crowded square what they still trust, and the students in the emptiest square to defend it.

## Notes on the core questions

**Core 1.** The setup removed steps 1 to 5 as people's work and shortened step 6. The lesson is that the slow part was before the writing. Students who say *"the AI wrote the memo faster"* have missed where the hours were.

**Core 2.** Good answers are concrete. Without the scheduled task, the firm learns about the returned check from the client. Without the drive, the task cannot read the file at 9:00 a.m. Without the skill's rule, the memo may call a returned check a timing item. Without the links, the second run makes four new pieces and leaves the old ones around.

**Core 3.** The rule: change the thing that makes the result. A fix by hand is right when the setup cannot be run again in time, or when the piece is the only one of its kind. Then the fix should be written down, and the setup changed afterward.

**Core 4 and 5.** These questions separate accountability from error prevention. A named approval says who answers for the figure. It does not find errors. On Tuesday the approval rested on a test that took the closing balance from Claude's run note instead of the bank's statement, so the error got through. Do not let students conclude that the morning check makes up for a weak test. Ask: which of Hamza's tests was designed wrongly, and what independent evidence should every reconciliation approval require? A good answer: the bank's own closing balance, from a file the setup never touched. Nisha knew what Hamza had checked, and how. Sana knew that, and she had read the memo herself. Neither could know that all 3,395 pairs were right. Ask what would raise their confidence: a larger sample, a second matching method, a comparison with the ledger's own balance.

**Core 6.** No answer is right. Point out that Tuesday's pack was READY, above the minimum, and wrong: *send if READY* would have sent it. But so did Nisha's approval. Ask what an approval adds, if it does not find every error: a name that answers for the figure, and a person who reads every correction before the client does. Hamza's reply, that the check protects the firm and not the approval, is a fair point and deserves a vote of its own. Students who choose *send always* should be asked who reads twenty packs afterward, and what happens on the day one is wrong. Students who choose *prepare only* should be asked how three people approve twenty packs a month. Students who charge for the hours should be asked who pays for the next setup. Students who keep $900.00 should say the sentence they would say to Walter.

## Takeaways for the wrap-up

1. Each part is useful alone. Joined, they do a whole job.
2. The time was saved where one part hands work to the next, with no copying in between.
3. Fix the setup, not the result.
4. Where a file lives decides which jobs can read it.
5. The setup changed the cost of the work. It did not change who answers for it.

## When students say "Google Docs does this too"

Take the point seriously (optional question 12). For writing and editing together, Google's editors are as good, and they keep a version history for slides, which Anthropic's pages do not describe for decks. Say so. Then ask where the firm's 17 hours were. They were in reading and matching the bank's lines, and in copying one result into several places. A document editor, with or without AI help, starts when someone has the figure. Do not let the class turn this into a contest between products: a firm could join other tools into a setup too, and ChatGPT has the same parts ([Exhibit 13](exhibits.md#exhibit-13-claude-and-chatgpt-side-by-side)). The lesson is the difference between a place to write and a setup that does the job.

## When students say "then the accountants are not needed"

Ask who found that slide 2 had no minimum on it, who knew that a bank's month is not finished on the first morning, and who wrote to Dana. Then ask what Hamza should be learning now, if not matching. The case's answer is in the right-hand column of the board.

## Running the class with non-accountants

The accounting is three lines and one subtraction. For accountants in the class, say what is simplified: a returned check is an adjustment to the books, the first download was made before the bank had finished posting September 30, and a real lender's test of cash is more detailed than one figure. Give non-accountants the two words they need: a timing item clears by itself, and an unrecorded item does not.

## When students ask about versioning the method

Optional question 15 and the *Going further* box in [What the setup did](what-the-setup-did.md#what-the-firm-had-now) raise it. The instructions changed on Tuesday and on Friday, and nothing recorded who changed them or who agreed. Let students see that this is the same problem as approving a pack: a change to the method needs a proposer, a reviewer and a record. Teams that keep instructions and skills as files can use version control and pull requests for this. Do not teach the tool here. [15.3 GitHub and the pull request](../../building-ai-workers/code-review/github-and-the-pull-request.md) teaches it later in the book.

## Notes on the series questions

**Question 16.** The setup is between a team AI Worker and a business AI Worker. It has shared rules, a lasting method and a schedule. It still depends on three named people, and its rules are mostly sentences.

**Question 17.** The rules (in the project), the method (in the skill and the instructions), the matched list and the run notes (in the drive), the four pieces (as artifacts), and the approvals (in Slack). The matching program itself was not kept, and did not need to be.

**Question 18.** *"Never send anything"* is backed by a setting: Nisha, as Owner, turned on only the shared drive as a connected app, so no chat in Claude can send mail. Note that connectors belong to each person's account, not to a project, so the lock is at the firm, not the project. The worker is set up separately: if its channel still holds the firm's mailbox, *"change nothing in the pack"* is only a sentence there. *"Never change a file you did not make"* is only a sentence, unless the drive connection is read-only for those files. *"Update these pieces, do not make new ones"* is only a sentence.

## What is assumed, not observed

The product facts come from Anthropic's help pages. The joined setup was not run for this case, so the run times and line counts are illustrative. Eight assumptions matter: that a chat in a project reads and writes a connected drive; that it runs the matching as a program in the cloud; that a run updates artifacts by their links; that three people can share the project, the skill and the pieces; that the worker in Slack can hold one more routine; that the worker can read the latest run note; that a sheet holds the month's lines; and that the lab's free path works as written. The [practical run sheet](appendix-practical-run-sheet.md) tests each of them.[^run-sheet] If your class finds that one behaves differently, that is a good discussion, and please tell the editors.

## Grading the lab (10 points)

Use the table in the [lab](lab-build-the-month-end.md#how-the-lab-is-graded). Two things to look for.

- In Part 4, the tests must be the student's own. An answer that says *"Claude confirmed it"* earns no points for that check.
- In Part 8, question 3, a student whose run made new pieces and who says so, with a fix, should get full marks. That is the lesson.

## The final assessment

The [final assessment](final-assessment-four-questions.md) covers the whole part, with a new organization. It is graded out of 10, and 7 is a pass. Its grading table is on the page.

## Facts and fiction in the case

| Fact, with a source | Fiction |
|---|---|
| What projects, skills, scheduled tasks, artifacts and the cloud computer do, and on which plans | The firm, the client, the bank and the people |
| That no piece reads another, and that a doc cannot leave a Team plan's organization | Every figure, time and line count |
| That a new session in the Code tab needs an artifact's link to update it | That this exact setup was run, and that a project chat updates artifacts the same way |

[^run-sheet]: The [practical run sheet](appendix-practical-run-sheet.md) says how to run the exercise and record it. It has not yet been run.