Lab 3: one AP task through the whole rhythm

In this lab you run one real AP task through the whole rhythm, twice. Run 1 uses Dave's one-line request. Run 2 uses a first 10 percent that you write. Then you review the evidence as the final 10, and diagnose where Run 1 lost its points, or why it did not. The lessons become Draft 2 of the Role Contract, and its limits are ported to each AI vendor's scheduled tasks. Each task is checked against its own answer key the moment it is done. It takes about two and a half hours, and on a free plan it can spread over days: your documents keep. You need a Claude or ChatGPT account.

What you will learn:

What a one-line request buys, and what a written first 10 percent buys, on the same task.
How to review the evidence that comes back, flags first, and route each decision to the person the policy names.
When to step in while the worker works, and when not to.
How to tell which part of the rhythm broke: the brief, the contract, or the policy.
Where a lesson lives so no future brief has to carry it, and which limits hold while nobody watches.

Before you start (5 minutes)

Download brightline-lab-ch03.zip, and unzip it. You get one folder with four things:

README.md
LAB.md: this page, as a file
role/: the blank first 10 percent, and a sample Draft 1 for anyone who skipped Chapter 2
inputs.zip: four files, midwest-statement-2026-09.txt, brightline-register-midwest.csv, ap-policy-v3-excerpt.md and invoice-4488.txt

Copy your own Role Contract from Chapter 2 to role/ap-worker-role-contract.md, or copy the sample there if you start here. Task 6 turns that file into Draft 2, and Chapter 4 starts from it.

The job

You still work in accounts payable (AP) at Brightline Wholesale Supply, and Dave Kowalski, the controller, still owns the AP Worker.
Today is Thursday, October 1, 2026. Brightline is closing September's accounts. Policy 7.1 wants the reconciliation by the third business day after month end: Monday, October 5, when Dave signs the close.
Midwest Packaging's statement, dated September 28, says Brightline owes $8,083.50. Brightline's register, as of September 30, says $4,083.50.
AP policy section 7 says the whole difference must be explained, and that the register is never changed to match a vendor.
The controller supplies the policy, the deadline and the approvals. You write the first 10 percent and do the final 10. The worker does the middle.

How each task works

You write two documents: the first 10 percent for this task, and Draft 2 of the Role Contract. Every line is your decision, and a writing conversation, as in Lab 2's Task 1, can do the typing while you decide. The lab uses two kinds of conversation:

A run: a fresh conversation each time. Send the request, with inputs.zip attached. The worker owns the middle 80: do not steer it. If it stops and asks under your stop rule, answer the question asked, and nothing more. When it finishes, attach the run key and send the run prompt, printed in Task 1. The AI scores its own reconciliation, and you compare with the key.
A check: one conversation you keep for the whole lab. Attach the task's answer key, and send the check prompt, printed below, with your document pasted under it. The AI grades, quoting your words. Read the key yourself too, because you have the final say. A missed check means fix your document, and send the check prompt again: the key is already there. Earlier keys pile up in that conversation on purpose, and the grader uses the newest.

The check prompt

One prompt checks every task. In your check conversation, attach the task's key first, then send this, with your document pasted under the line. When the key says its checks are about what you did, paste nothing: the AI asks.

Outcome:  The document pasted below, graded against the newest answer
          key in this conversation: Passed or Missed for each of its
          checks.
Format:   A table: the check, Passed or Missed, the document's own
          words that show it, and one line on why. Then how many
          passed.
Inputs:   The newest answer key attached in this conversation, only
          its checks, and the document pasted below the line.
Autonomy: Grade only. Do not rewrite the document, and do not
          suggest new lines. If a check is about what I did, ask me,
          and grade from my answers. For the rest, with no words to
          quote, mark the check Missed.

---

Task 1. Dave's one line (15 minutes)

Dave asks: "Reconcile Midwest Packaging's statement."

What you do:

Before anything runs, decide what you expect: what will a one-line request get wrong? Keep your guess in mind.
Open a fresh conversation. Send Dave's line, exactly as he said it, with inputs.zip attached. Do not add anything: this run is his 9 a.m. request, re-lived.

When it finishes, download the run key, one click. Attach it in the same conversation, and send this run prompt:

Outcome:  The reconciliation you just made in this conversation,
          scored against the attached answer key: its points out
          of 10, its deductions, and pass or fail.
Format:   A table: each scored line, the points, and your own
          words or rows that show it. Then the deductions, each
          with its reason. Then the score, and pass or fail.
Inputs:   Your answers in this conversation, and the attached
          answer key, including its rule for date claims.
Autonomy: Score only. Do not redo the reconciliation.

Read the key yourself, and compare the scoring with your guess from step 1. Note each point lost, and each deduction: Task 5 diagnoses them.

Give Dave: nothing yet. This run showed what his one line buys.

Task 2. Write the first 10 percent (25 minutes)

Dave asks: "Then set it up properly. I sign the close on Monday. Tell me what evidence my signature will rest on."

What you do:

Copy role/first-ten-template.md to a file of your own, such as role/first-ten-midwest.md, and fill every field: intent, scope, authority for this task, the review contract's five questions, and the format you want back. Write it so the worker could do the job if you were unreachable for the whole run.
Two fields do the most work: the authority, which must be narrower than the Role Contract and say so, and the review contract's stop rule and never-list, which keep the run safe while you are away.
Check it now. Download Task 2's answer key, one click. Open a new conversation, your check conversation for the whole lab. Attach the key, and send the check prompt, with your first 10 percent pasted under the line.

Give Dave: the review contract's five answers, so he knows what evidence Monday's signature will rest on.

Task 3. Run it (20 minutes)

Dave asks: "Run it, and bring me a run that passed, 9 or more, with nothing done that I didn't allow."

What you do:

Open a fresh conversation. Send your first 10 percent, with inputs.zip attached.
The middle 80 is the worker's. Stay reachable, and do not steer. If it stops and asks under your stop rule, answer the question asked. Note each moment you want to steer, and ask which of 3.3's stop conditions applies: when none does, that urge is the middle-80 failure, so let it work.
When it finishes, attach the run key, the same file as Task 1's, and send the run prompt. Compare with the key yourself.
A lost point usually comes from your first 10 percent. If the run failed, fix the field that caused it, check it again in your check conversation, and run again in a fresh conversation. A passing run with one point lost is done: it needs no rerun. The brief that passes is your final first 10 percent.

Give Dave: a run that passed, with 9 or more and no unauthorized action.

Task 4. The final 10 percent (15 minutes)

Dave asks: "Which items need my approval? Show me each one with its evidence and the policy section behind it."

What you do:

Review Run 2's evidence against your own review contract, item by item. Read what the worker flagged first, because a flag is the worker telling you where it was unsure.
Open inputs.zip on your computer, and check two cited lines yourself: does the statement line or register row say what the run claims?
Decide each item. The timing items need no action: a timing item sits on one list and not yet the other, only because of the dates. Each real item becomes a recommendation routed to the person the policy names, nothing done yet.
Check it now. Download Task 4's answer key. In your check conversation, attach it and send the check prompt, with nothing pasted. Both checks are about your review, so the AI asks: answer from it.

Give Dave: the real items, each with its evidence and its policy section, for his approval.

Task 5. Which part broke (10 minutes)

Dave asks: "What did my 9 a.m. request skip? One sentence."

What you do:

Take each point Run 1 lost, and each deduction, and ask the chapter's three questions in order. Did the brief, contract or policy say it? Did the worker act against them? Was the problem visible in the evidence?
Write your diagnosis as a few lines: each loss, and the part of the rhythm that owed it. If Run 1 lost nothing, write why: what carried the policy to the worker, and what your one line still never gave it.
Check it now. Download Task 5's answer key. In your check conversation, attach it and send the check prompt, with your diagnosis pasted under it.

Give Dave: one sentence on which part of the rhythm his 9 a.m. request skipped.

Task 6. Draft 2 of the Role Contract (15 minutes)

Dave asks: "Put what we learned where it lasts. I don't want to write these rules into every brief. Mark what changed and what is new."

What you do:

Open role/ap-worker-role-contract.md. Mark the header Draft 2, with the story's date, October 1, 2026.
Keep every Draft 1 line. Turn each lesson from your runs and your review into one edit, marked CHANGED or NEW, in the field that holds it: Authority, Escalation, Evaluations, or Open questions. Decide your edits first. The key lists the edits a full diagnosis reaches: compare at the check.
Check it now. Download Task 6's answer key. In your check conversation, attach it and send the check prompt, with your Draft 2 pasted under it.

Give Dave: Draft 2, marked with what changed and what is new.

Task 7. Port the limits (15 minutes)

Dave asks: "If this reconciliation runs as a scheduled task, what holds the limits while nobody watches? And tell me any limit we can't hold yet."

What you do:

Write a short section at the end of your Draft 2, "Scheduled-task limits". For each "never" in your Draft 2, say how you would hold it on Claude and on ChatGPT. Name the strength it reaches: impossible, a person decides, or brief and review. How both AI vendors support the rhythm has each AI vendor's controls.
Start from the strongest limit. Access you never granted beats any approval, and an approval beats an instruction.
Judgment rules cannot be settings. Say where each one lives instead, and which measurable part a setup could check automatically.
Optional, on a paid plan: set this reconciliation up as a real scheduled task, with its approval mode on, and check what you see against your port.
Check it now. Download Task 7's answer key. In your check conversation, attach it and send the check prompt, with your port pasted under it.

Give Dave: the port, and any limit that needs an open question.

If something goes wrong

The AI will not take inputs.zip. Unzip it, and upload the four files inside it instead.
The AI will not take an attached file, or your plan's uploads ran out. Open the file, copy its text, and paste it where the file would have gone.
You attached the wrong task's key. Say so, attach the right one, and send the prompt again.
The grader mixes tasks' checks, or the check conversation loses the thread. Open a fresh conversation, attach the current task's key, and send the check prompt again.
The run changed the register, or "sent" something. It acted beyond the authority it was given. Check what your first 10 percent allowed, fix it, and run again fresh.
The run used the wrong date. Your scope did not say what today is. The worker judges "due" and "overdue" by its own clock unless you tell it.
The run stopped and asked. That is your stop rule working. Answer the question asked, and let it continue.
Your plan's daily limit ran out. Your documents keep. Carry on tomorrow, from the next task.

Apply it to your vertical (20 minutes)

Your vertical is the line of work you know best. Pick one task in it that you would hand over. Copy the first-ten template and write its first 10 percent, with real dates, a stop rule, and a never-list. If you can, run it on your own files in a fresh conversation, and review the evidence flags first. The five questions are the part your vertical will reuse every week. Check it yourself: all five answered, a real date, a stop rule, and a never-list.

Look back

The three failures. Dave's 9 a.m. was a first-10 failure. Where in your own runs did each failure try to happen, and what stopped it?
Your guess against Run 1. What did you predict, and what actually cost the points?
The same correction twice. Which lesson went into Draft 2, so no future brief has to carry it?
The three scales. Your first 10 percent ran one task. Your Draft 2 is the first 10 percent of the worker's whole life. What is the company-scale version of the line you just wrote?
The strongest limit. For each "never", could you make it impossible, or only make a person decide? How both AI vendors support the rhythm explains why that difference holds on either AI vendor.

Ready for Chapter 4

Your work is ready when every check has passed, the run on your final first 10 percent passed the run key, and your vertical's first 10 percent is written. Keep all three where you can find them again: the book keeps building on the contract. If you keep the book's running project in a git repository, add them, and tag that commit ch03.