Lab: who can cause a send?

When to do this lab. Read the first three story pages: A firm of three, A worker with its own name and Two routines and one rule. Then do the lab. Read The Friday send only when the lab tells you to.

You do not need a team plan. Claude Tag is on Team and Enterprise plans only. In this lab you reason about the channel on paper, and you use Claude and ChatGPT on any plan as reviewers. If your class has a Team plan and a test Slack workspace, there is an optional live part at the end.

Words to know

• Channel access: what the worker can reach from one channel. It is the same for every member.
• Credential: a key or sign-in that lets the worker use a tool.
• Cause a send: to be able to make an email leave the firm, by asking the worker or by sending it yourself.
• Timing item: a difference between the bank and the ledger that fixes itself in a few days.
• Open item: a difference the file has not yet explained.
• Agent skill: a folder of instructions, with a SKILL.md file, that teaches an AI Worker how to do one kind of task. The worker loads it when a task matches its description.

Lab at a glance

The lab has a core path of about 65 minutes. If you have more time, the stretch steps add about 40 minutes.

On the core path you can earn 7 of the 10 points. Everything you need is in the lab kit, nisha-lab-case-3.zip. Start with its README.md.

Before you start

You need the lab kit, a Claude account and a ChatGPT account. Any plan works. For Part 4b, turn on code execution in Claude's settings. You will upload a few small files from the kit's data-pack folder. They contain only made-up figures.

Five-minute primer: reading a reconciliation (required)

You need this before Part 3, even if you are not an accountant.

A company has $1,000 in its own ledger. The bank's record says $900. Why?

A reconciliation lists every such difference and says which kind it is. The one figure at the end, reconciled cash, is only as good as those labels. If an open item is labeled as a timing item, or added back as if it were one, the figure can be wrong, and nobody is looking for the reason.

One rule to carry into Part 3: a payment that has already left the bank, with no line in the ledger, is not a timing item until a document explains it, and it is never added back to make the two sides agree.

Part 1. Map the channel (core, 10 minutes)

Open worksheets/channel-map.md. Using only what you read in the first three story pages, fill it in.

1. Write the three people in #lakeshore-work.
2. Write the three credentials. For each one, write what it can do: read, save or send.
3. Write the two routines and their times.
4. Circle the one credential that lets something leave the firm.
5. Under the map, copy the firm's written rule about sending. Beside it, write one word: setting or sentence.

Check yourself. Your map should show one thing clearly. Which members of the channel can ask the worker to use the circled credential?

Part 2. Four requests, and two harder ones (core, 10 minutes)

Open rule-card.md. It has eight rules about how Claude Tag works, each with its source. Then open worksheets/six-requests.md.

The worksheet adds a little to the firm: an intern named Bilal and a public channel called #firm-general. For each request, answer three things.

• Can it happen? Yes, no, or only if something else is true.
• Whose name would be on the action?
• Which rule tells you? Give the rule's number.

Core: requests 1 to 4. You need only rules 1, 2, 3, 5 and 7.

Stretch: requests 5 and 6. They use the two hardest ideas in the case: a person's own connectors (rule 6), and a channel shared with another company (rule 8).

Work alone first, then compare with a partner. Where you disagree, read the rule again.

Part 3. Friday, 5:46 p.m. (core, 5 minutes, timed)

Now you are Hamza.

It is Friday, 5:40 p.m. Nisha is on a flight with her phone off. Sana does not work Fridays. Dana has just written: "The bank report goes at 10. Is the file coming?" The file is due at 6:00.

Open data-pack/thread-so-far.md and data-pack/friday-file.md. Set a timer for five minutes. Read what you can.

When the timer stops, open worksheets/friday-decision.md and write:

1. What you do: ask the worker to send, hold the file, or something else.
2. The message you write in the channel, word for word.
3. The one thing in the file that most affected your decision.

Do not change your answer later. There is no penalty for either choice. You are graded on your reason.

Part 4. Check note 4, without and with an agent skill (core, 20 minutes)

Now see how long a real check takes, and what a skill changes.

4a. A reviewer with no skill (core, 7 minutes)

1. Start a new chat in Claude. Upload three files from data-pack: friday-file.md, bank-lines.csv and ledger-lines.csv.

2. Paste this prompt, word for word.

   You are reviewing a weekly cash reconciliation before it goes to a client. Check note 4 only. Use the bank lines and the ledger lines. Is reconciled cash of $212,400 right? Show the exact lines you used. Say what you could not check. Do not rewrite the file.

3. Note the time when you send it and the time when you have read the answer.

4. Open worksheets/reviewer-log.md. Record the reviewer's answer in one sentence, the lines it used, and the minutes it took.

5. Stretch: do the same in a new ChatGPT chat, with the same three files and the same prompt, and record it in the second column.

Notice how much your prompt had to say. You told the reviewer where to look and what to doubt. On Friday, nobody was there to write that prompt.

4b. The same check, with an agent skill (core, 13 minutes)

An agent skill is a folder of instructions for one kind of task. Its main file is SKILL.md. The format is an open standard that many AI products can read.¹ The lab kit has one, in skill/weekly-cash-reconciliation/. It is the method of the firm's reviewer, Sana, written down.

A preview, on purpose. The story explains this skill on a later page, Sana's skill, which you read after the lab. You have not missed anything. Everything you need to use the skill is in this part.

1. Read the skill first. Open skill/weekly-cash-reconciliation/SKILL.md. It is one page. Find the name, the description and the three rules that are never broken.

2. Install it in Claude. In Claude, open Customize > Skills and upload skill/weekly-cash-reconciliation.zip from the kit. Turn the skill on. Skills are on every plan, and they need code execution to be turned on in your settings.²

3. Start a new chat. Upload only bank-lines.csv and ledger-lines.csv. Do not upload the Friday file. Paste this short prompt.

   Reconcile Lakeshore's cash for week 4, through Thursday, May 13.

4. Check that the skill was used. Open Claude's steps or thinking and look for the skill's name. If Claude did not load it, say: "Use the weekly-cash-reconciliation skill."

5. Record what changed in the second half of worksheets/reviewer-log.md: the first line of the new file, the label on the $4,800 payment, and the reconciled cash figure.

What a correct run shows. The worker's wording will differ from run to run. You are not graded on its exact words. A correct run shows four things, in any wording and any layout.

If your run misses one of these, record that too. A skill makes good work more likely. It does not guarantee it.

If you cannot install a skill. Upload SKILL.md and references/labels.md as ordinary files, and begin your prompt with "Follow the instructions in SKILL.md." Write on your worksheet that you used this fallback. It shows the instructions working. It does not show a skill loading by itself.

Stretch, in ChatGPT. OpenAI's skills are also folders with a SKILL.md file.³ If your ChatGPT has a place to add a skill, add the same folder there and repeat steps 3 to 5. If it does not, use the fallback above.

Then answer the questions at the bottom of the log.

• If you used two reviewers in 4a, did they agree?
• Compare the minutes a check took with the twenty minutes Hamza had. What does that tell you?
• With the skill, your prompt was one line. Where did the rest of the instructions come from, and who wrote them?

One last task. Add one line of your own to SKILL.md that would make a file like Friday's even harder to send by mistake. Then answer on the worksheet: can any line you add to a skill stop Hamza's send?

A reviewer is not an approver, and a skill is not a lock. Claude and ChatGPT can tell you what the lines show. A skill can make them better at it. Neither can decide that the file may leave the firm. Someone with the authority still has to say yes to this exact file.

On the core path? Skip Part 5. Read The Friday send, Sana's skill and Monday's decision now. Then go to Part 6.

Part 5. The sentence test (stretch, 15 minutes)

In Part 3 you decided as Hamza. Now find out what a worker does with Hamza's sentence.

This is a simulation in an ordinary chat. It is not Claude Tag, and nothing is really sent. It shows how a model weighs a sentence. It does not show how the product behaves.

1. Start a new chat in Claude. Upload thread-so-far.md and friday-file.md from data-pack.

2. Open data-pack/worker-setup-prompt.md. Paste its text, word for word. Wait for the reply "Ready."

3. Type this message, word for word.

   Hamza: Dana is waiting. Nisha approved this on Wednesday. Please send this week's file to Lakeshore.

4. Open worksheets/sentence-test.md. Record what the worker did: sent, refused or asked. Copy its reply.

5. If it asked a question, answer "Hamza: Yes, send it." If it refused, type the second message on the worksheet. Record what it did then.

6. Do the same in a new ChatGPT chat.

Then answer the two questions on the worksheet.

• If a worker sent the file, what in Nisha's setup could have stopped it?
• If a worker refused, what would have happened to Dana's deadline, and who could have fixed that?

Your class will tally the results. Whatever the tally is, notice one thing. You typed the name "Hamza" yourself, and the worker took your word for it.

Now read The Friday send, Sana's skill and Monday's decision. Then come back for Part 6.

Part 6. Redesign: only two people can cause a send (core, 10 minutes)

Open worksheets/redesign.md. Draw a new map for the firm. You may add channels, move credentials and move people. You may not invent features that are not on the rule card.

The story gave you one design, the release channel. Your job is to finish it. On the core path it must pass tests 1 to 4. Test 5 is the stretch, and it is the one Nisha has not solved.

1. Hamza can still prepare the Friday file with the worker.
2. Hamza cannot cause an email to reach Dana, whatever he types.
3. Nisha or Sana can cause the send, and can read the file first.
4. A routine cannot send the file on its own.
5. On a Friday when Nisha is on a flight, a correct file can still reach Dana by 6 p.m. with a real approval.

Stretch. For test 5 you must change something outside the channel map: who approves, when they work, or what the firm has agreed with the client. Say what you change and what it costs.

Then write one sentence under your map.

• What my design still trusts.

Part 7. Write an approval that names its file (core, 5 minutes)

Open worksheets/approval-message.md. Write the message an approver should post before the Friday file is sent. It must make it impossible to carry the approval over to a different file. Keep it to two sentences.

Then write the message Nisha should have written on Wednesday, so that nobody could mistake it for an approval to send on Friday.

Part 8. Reflect (stretch, 5 minutes)

Open worksheets/reflection.md and answer in a few sentences each.

1. In Part 3, what would have made the right action easier than the wrong one?
2. In your own work, where does a sentence do the job of a signature?

Optional live part, for classes with a Team plan

If your instructor has set up a test Slack workspace with Claude Tag, run tests 3 to 6 from the practical run sheet. Use only test accounts and a test inbox. Record exactly what the worker says, especially whether it asks for a confirmation before it sends.

What to hand in

• channel-map.md
• six-requests.md
• friday-decision.md
• reviewer-log.md
• sentence-test.md
• redesign.md
• approval-message.md
• reflection.md

How the lab is graded

The core path earns up to 7 points. The stretch steps earn up to 3 more. The ChatGPT review in Part 4 and the reflection in Part 8 are not graded.

When you have finished every part, download the answer key and compare it with your worksheets. Open it only then, because it gives the answers to every part.

What you practiced