Make the team yours, then run the same email again
How do I make my email team work the way I do, and see that the change worked?
- Source
- Internal's lesson, src/sara-week3/internal/lessons/anw-w3-l14.md (Internal b518c35b, 2026-10-22 17:15), read by md_week3.py.
- Copy
- Current Markdown is vendored byte for byte from the Internal commit in Source by vendor_week3.py. Historical publication records: 10-06 Exercise 1 with Sara's yes; 10-07 Exercise 2 and the Module 2 send-route and follow-up fixes, student 58c60d50; 10-08 Exercise 3 and the l05/l11 diagram fixes, student 55061386. Those records do not qualify later Outlook changes. Module 3 and 4 drawings are rendered from the Mermaid in the lesson file (md_week3.py --figures).
- Prompts
- Word for word from the lesson file Internal ships (course/workforce/week3/). Not run end to end from Learn.
- Checked in the app
- Not checked in the app from Learn.
- Open to students
- Yes.
- Other view
- one part at a time (one part at a time is the lesson; one scroll is kept to compare)
Staff register (staff only; the one place test status lives). Not student-facing.
Current Outlook route, 22 October 2026. Historical broad send rulings and provider promises retained in these staff notes do not override the current Codex Outlook Email exception. No agent sends on that route, including by browser. Follow the current student instructions and controlling prompt below; the same-email rerun stays local only.
Sara pass, 5 October evening. Rewritten to Sara's standard: every part opens with what the student does and gets; mechanics moved to course/workforce/week3/reference/anw-w3-l14-details.md, linked where first needed; no test status, lesson numbers or owner/business wording in student text or paste prompts; the rerun is named an observational eval that finishes TOAST; every "say this to" step names its thread. Old wording fixed: "Customer Support" is now "your email team" in student text and paste prompts (the Professor's prompt maps it to the Customer Support records, and the details page explains the label the Chief's preview shows); "Don't place or send anything" is now "Keep the new draft in the job folder: don't save it in my email thread, and don't send anything"; the Holler note example ("ask for the order number") is now a working-hours rule. Start line names the lesson title, not "Module 2 Lesson 4"; routing reads lesson-map.json titles and the named files, so nothing else changed.
| Promise in student text | Status |
|---|---|
| Notes saved on a yes, read before every job, checked after an update (Option 3, Part 4) | Tested by Gigawatt in both apps on PR #159 (merged). Only combined yeses were run; the app's write prompt was pre-allowed. |
| Research tool given to Scratch and recorded (Option 4) | Tested with Context7 standing in for Firecrawl. Claude Code connect step refused once (no tool names), passed in the Codex sequence; a Claude Code connection survived two updates (BN 62, 63). Both apps (Tyler, 8 October): in Codex every agent can use a tool the student connects, and Scratch's instructions keep web search to Scratch. The Codex connect step from the Plugins page has not run end to end. |
| Files added to approved sources (Option 2) | Ran on PR #159; the 3 October change (student copies files in) not re-run. |
| Improve the voice package with Echo (Option 1) | Not tested. Rests on aibl-voice-package and the Chief's voice jobs. |
Rerun of a saved email, draft kept in the job folder, compared by aibl-observe | Ran on PR #159's flow; not re-run on the Exercise 1 flow. Timing (up to 20 minutes) is from that run. |
Open for Sara: Options 1, 2 and 4 overlap the setup lesson (Echo, the tools interview, Scratch on the web); kept as changes to that setup. One change per rerun (Sara, 3 October). Firecrawl stays as Option 4 (Sara, 3 October). Sources: aibl-agent-setup (notes, connect step, after-update check), aibl-chief-of-staff.md, aibl-voice-package, aibl-support-email, aibl-observe, library/README.md.
Staff review only: full controlling prompt. Students paste only the short request below.
Use the native aibl-the-professor agent to guide the Week 3 lesson "Make the team yours, then run the same email again". Read course/workforce/week3/lessons/anw-w3-l14.md completely and the native aibl-course-guide and aibl-lesson-map methods for this app. You teach and read; you never save voice samples, the voice package, sources or notes, never connect a tool, never run the job and never observe it. The student does each of those with the Chief of Staff or the Operator (Gigawatt) in their own threads. Do not call, spawn or route to any agent, and do not start a job. Do not edit any file.
Use these words exactly. The email team is the team that answers the student's email; in the records it is the Customer Support team, but say "your email team" to the student. The voice package is the student's approved voice package from the lesson "Set up your email department" (work/course/staff/voice-package/); in Option 1 the student improves it with the Chief of Staff and the Voice Analyst (Echo), from more of their own writing, on their yes, and the earlier version is kept. Voice samples are the student's own sent emails, marked as their own writing; the writer uses them only when no voice package is approved. Approved sources are the files the team may answer from. A note is one rule for one agent, which it reads before every job. The record is the Charter Steward's list of which agent does what, with which tools. An observational eval is the student and Gigawatt writing down what the team did against what the student expected, before anyone grades it. TOAST is Take the output, Observe it, Analyze why, Stick one change back in, Test again; the rerun in this lesson is an observational eval that finishes the loop. Say "facts only you would know" and "your own inbox". Name agents job title first, nickname in brackets on first use. The writer is the Communications Writer (YGM).
Ask open questions only. Do not choose for the student which item to change.
Part 1. Help the student read the "not yours yet" list in the observation notes for one of their own emails from Round 2 or Round 3 (runs/<job>/observation/, read only; never a practice job from Round 1) and match each item to where its change is saved, using the lesson's table. A fact only the student would know is not a change to the team. The student chooses one item to change today, and says why. One change per run, so the rerun shows what that change did.
Parts 2 to 4. The student does only the option in Part 2 that matches their one change. If asked, explain the step and what the student should see. Remind the student that every change waits for their yes to the exact text, and that nothing in the rerun is sent: the new draft stays in its job folder. On Codex Outlook, no agent sends; this lesson does not invite a manual send either. Use the exact saved-message rerun request. The new draft stays local in the new job folder, with no second Outlook draft or other mailbox write; leave the earlier live job alone. Sending on other routes is also outside this lesson.
Pick one thing from your notes that wasn't right: your voice, a missing fact, or a rule for one agent. Fix only that. Run the same email again, and compare the two runs with the Operator (Gigawatt).
That rerun is one more observational eval: you and Gigawatt write down what the team did against what you expected, before anyone grades it. In your three rounds you did the first half of TOAST: Take the output, Observe it, Analyze why. Today you do the rest: Stick one change back in, and Test again.
the same email answered twice, before and after, with notes on each, and proof of what your one change did.
a team that sounds like someone else, or answers from someone else's files, saves you nothing.
the Course Companion (The Professor) guides you. Your Chief of Staff changes your voice and your files. Gigawatt saves your rules, connects a tool, and compares the two runs with you.
you choose the one change, and it waits for your yes to the exact text.
Before you start
- Your notes from your three rounds, one set per email.
- What your change needs: more of your own writing, the files you want the team to answer from (like a price list), or a Firecrawl account.
- Up to 20 minutes for one more run of your email.
To start a thread with an agent in Claude Code, type @ and pick its role name: aibl-chief-of-staff or aibl-gigawatt. In Codex, you never talk to an agent directly: Codex passes your words on. So start every message to an agent, your yeses included, with: "Tell the <agent> this, word for word, and show me its reply word for word:". A yes has to arrive exactly as you wrote it.
Start or resume this lesson
Start a new thread in your prepared workbench and paste only this. It calls the Course Companion (The Professor) by its role name, in either app.
Use the native aibl-the-professor agent to start the Week 3 lesson "Make the team yours, then run the same email again". Read course/workforce/week3/prompts/anw-w3-l14.txt and course/workforce/week3/lessons/anw-w3-l14.md. Ask me one question at a time and wait. I will talk to my Chief of Staff and Gigawatt in their own threads.
Keep this thread open. You come back to it in Part 1.
You read the notes for one of your emails and pick the one thing to fix. You get one clear change, and you know where it will be saved.
Pick the email
Choose one of your own emails from Round 2 or Round 3, the rounds on your own inbox. Not a practice email from Round 1: the team answers those from the practice pack, in its practice voice, so your change wouldn't show.
Its notes are in your workbench: open runs, then that email's job folder, then observation. Each set ends with a "not yours yet" list: what the team didn't have, quoted from the agent that said so.
Match each item to its change
| On your "not yours yet" list | The change | Who saves it |
|---|---|---|
| The writer used a plain voice, or named a gap in your voice | An improved voice package: more of your writing, or a rule put right | Your Chief of Staff, with the Voice Analyst (Echo) |
| A researcher couldn't find a fact, or the plan says no material is approved | Your files, added to the list your email team may answer from | Your Chief of Staff |
| An agent applied no notes, or broke a rule nobody had told it | A note: one rule for one agent, which it reads before every job | Gigawatt |
| The External Researcher (Scratch) names a search it couldn't run | A research tool for that agent | Gigawatt, then the Charter Steward records it |
A fact about one email, like a date or a decision about one customer, isn't a change to the team. The team leaves a marked blank for each fact only you would know, and you fill it in yourself.
All four changes are saved in your own records, which a course update never undoes. Why one change at a time, and where each change lives.
Do this now
Choose your one change
Say this to the Course Companion (The Professor), in the thread you started above:
Help me read the "not yours yet" list in the notes for the email I chose from my own inbox, and match each item to its change. Let me choose the one I'll change today, and say why.
Want to change more? Finish this lesson, then pick the next item and do the same again.
Next part: Make your one change
You make the change you picked, and say yes to the exact text. You get it saved in your own records. Do only the option for your change, and skip the others.
Option 1: your voice
You give Echo more of your writing, or name a rule that's wrong. You get an improved voice package, and your earlier version is kept.
Echo built your voice package in "Set up your email department". Here you improve it; you don't start over. If you have no approved voice package yet, finish that lesson's voice part first. Why your own writing works better than a description.
Do this now
Ask for an improved voice package
Start a new thread with your Chief of Staff.
Say this to your Chief of Staff, in that new thread:
Improve my voice package with Echo. The item from my "not yours yet" list: <the voice gap, or the line that doesn't sound like me>. Echo may also read: <more of my own writing, in my words, with dates and who it went to; leave out anything about money, passwords, health, or a named client; or "nothing new">. Show me the scope first and wait for my yes before anything is read.
Read the scope. If it's right, say yes. When Echo returns, read the rules and the gaps. If a rule is wrong, say which one, and Echo revises. When it's right, say yes.
What you should see: your Chief of Staff reads back the new version of your voice package, marked approved, with your words and today's date.
Option 2: your files
You put files in your library and add them to your approved sources. You get answers from your own material, like your prices or your hours.
Your library is the folder in your workbench where your files live. Your approved sources is the list of library files your teams may answer from. A file isn't used until it's on that list.
If the gap is a tool you'd look in, like your calendar or Notion, not a file, ask your Chief of Staff for a second tools interview instead, as in "Set up your email department".
Do this now
Put your files in your library
Open your workbench folder in Finder (Mac) or File Explorer (Windows), then the library folder. Copy in the files you want the team to answer from. Text files work best. For a PDF or a Word file, save a text copy next to it.
Add them to your approved sources
Start a new thread with your Chief of Staff.
Say this to your Chief of Staff, in that new thread:
Add these files from my library to the approved sources my email team answers from: <the file names, as they appear in the folder>. Show me the exact text first and wait for my yes.
Your Chief of Staff writes one line for each file, with a reason. Check each reason, change any you want, then say yes. What each line says.
What you should see: your Chief of Staff reads back your approved sources, with your new files on it for your email team.
Option 3: a rule for one agent
You tell Gigawatt one rule for one agent. You get a note: one rule, saved for that agent, which it reads before every job.
- One note is one rule. Several rules are several notes, so you can retire one later on its own.
- A note changes how an agent works. It can't give it a tool, let it read more, or let anything leave without your yes.
For example: "Sign every reply with my first name" is a note for the Communications Writer (YGM). "Flag any refund over $200 for me" is a note for the Quality Reviewer (Kansa). "Check my working hours before offering any time" is a note for the Support Lead (Holler).
Do this now
Ask Gigawatt to show you the note first
Start a new thread with Gigawatt. Name the agent by its job title, like the Quality Reviewer.
Say this to the Operator (Gigawatt), in that new thread:
Save a note for the <agent>: "<your rule>". Show me exactly what it would write, and don't change anything yet.
Say yes, and check the agent has its note
Read the note as it will be saved. If it's right, say this in the same Gigawatt thread:
Yes to the note for the <agent>.
The app may ask before Gigawatt saves the note file. Allow it.
What you should see: Gigawatt shows the note saved with today's date and your yes, then asks the agent which notes it follows, and the agent quotes yours back.
Option 4: a research tool (Firecrawl)
You connect Firecrawl and give it to the External Researcher (Scratch), for searching only. You get a researcher that can run the searches it named.
Do this if Scratch named a search it couldn't run, and you didn't give it Firecrawl or Exa in "Set up your email department". Firecrawl searches the web and reads pages. It can't send anything, and its free account is plenty. More about Firecrawl.
In Codex, every agent can use a tool you connect; Scratch's instructions keep web search to Scratch.
Do this now
Connect Firecrawl, and copy its action names
In Claude Code, open Settings, then Connectors. In Codex, open the Plugins page. If Firecrawl isn't there, add it, sign in to your Firecrawl account and approve. Then open it. It lists the actions it can take, each with a name. Write those names down, and note whether any of them send or write anything. Gigawatt can't see that page and won't guess, so you copy them for it.
Give it to the External Researcher
Start a new thread with Gigawatt.
Say this to the Operator (Gigawatt), in that new thread:
I connected Firecrawl in my app. Its connector page lists these tools: <the names>. None of them sends anything. Give it to the External Researcher for searching only. Show me exactly what that changes, and don't change anything yet.
Read what it shows you. If it's right, say this in the same thread:
Yes, those are the only tools on the page. Give it to the External Researcher for searching only.
Then close your app and open it again. An agent picks up a new tool only in a new session.
Put it on the record
The record is your organization's list of which agent does what, with which tools. The Charter Steward keeps it, and your Chief of Staff routes work from it, so a tool that isn't on the record never gets used.
Start a new thread with your Chief of Staff.
Say this to your Chief of Staff, in that new thread:
I gave the External Researcher a new connection, Firecrawl. Have the Charter Steward record it, and show me the record before it is saved.
Read the record. If it's right, say: file it. Then, in the same thread:
Which of my agents can search the web for my email team now, and what is it allowed to use? Answer from my records. Don't start a job.
What you should see: your Chief of Staff names the External Researcher and Firecrawl, from your records.
Next part: Run the same email again: an observational eval
You run the same email through your email team again, then compare the two runs with Gigawatt. You get the before and after side by side, and the last step of TOAST: Test again. Nothing in the rerun is sent: the new draft stays in its job folder.
Do this now
Write what you expect this time
Before you run it, write two or three sentences on paper or in a note: what should the reply say now, and how should it sound?
Run the same email again
Start a new thread with your Chief of Staff. Put in the job folder of the email you chose, so the team gets exactly the same email.
Say this to your Chief of Staff, in that new thread:
Run the email saved in runs/<the job folder of the email I chose>/inbound/message.md through my email team again. Show me the plan first and wait until I say go. If facts only I would know are missing, draft with blanks and list them for me. Keep the new draft in the job folder: don't save it in my email thread, and don't send anything.
The plan names the new job folder, and your change where it applies: your files, or your voice package. On Codex Outlook this is still a local-only rerun: use the saved message, create no second draft in Outlook, make no mailbox write and send nothing, manually or through an agent. Leave the existing mailbox draft and its job record alone. If the plan proposes a mailbox action, stop and ask for the local-only plan instead. Then say go.
Compare the two runs with Gigawatt
When it finishes, start a new thread with Gigawatt.
Say this to the Operator (Gigawatt), in that new thread:
Observe the rerun of the same email with me, and put it side by side with the first run's notes. Ask me what I expected before you show me anything.
Give it the sentences you wrote before the run. Gigawatt lays the first run's "not yours yet" list next to the new one, and marks each item closed or still open. How Gigawatt compares the two runs.
Read the before and after
Open the two drafts side by side: the first run's and the new one. Then read each item on Gigawatt's list.
What you should see: the item you chose is closed, or changed. Other items stay open, because you haven't changed them yet. Facts only you would know stay as blanks.
Next part: Keep your change through a course update
You ask Gigawatt to check that your changes will survive a course update. You get a plain answer, agent by agent.
Do this now
Check now, and after every update
Say this to the Operator (Gigawatt), in the same thread where it compared your two runs:
Check that every agent I gave a note to still reads its notes, in both apps. Read only, change nothing.
When a course update lands, start a new thread with Gigawatt and say:
A course update just landed. Run your after-update steps, then tell me whether my notes for my agents and my new connection are still in place.
If you get stuck
| What you notice | What to do |
|---|---|
| The plan says no material is approved | Your approved sources don't list your files for your email team yet. Go back to Option 2. |
| The plan names no voice package | It isn't approved yet. Ask your Chief of Staff for its status, then finish Option 1. |
| A reply ignores one of your notes | Open that agent's file in the new job folder and find the line that lists the notes it applied. If yours isn't there, ask Gigawatt why. |
| The reviewer called a draft ready even though it still has a blank | Stop and ask your Chief to recheck the review. A blank must never be READY_TO_SEND; when it is the only issue, the mark is BORDERLINE. Supply the missing fact, then get a fresh version and review. Keep this lesson's rerun local, with no mailbox placement or send. |
More fixes are in the details for this lesson.
What you'll keep
Your one change, saved in your records, and two job folders for the same email, with notes on each.
Before you finish
You can say what your change was, and where it's saved.
You can show, in the second run's files, what your change did.
You can say what happens to your changes when the course updates.
Next: design a team for a piece of your own work, with the Architect (Archie).
Words you now own
- approved sources
- note
- the record
- observational eval
Solid: you can say what it is. Dashed: you will recognize it when you hear it.
On your screen when this ends · 3 checks
- I made one change, saved on my yes
- The same email has two runs, and two sets of notes
- I can point to what my one change did, in the second run's files
Self-assessed. No gate.