Set up your email department: the details

This page goes with the lesson "Set up your email department". You do not need it to finish the lesson. Read a section when the lesson links to it. The "What the course has tested in your app" table on the page of the lesson "Run three observational evals, then review and send" is the one place that says what is on in your app; nothing here turns anything on.

Your email team

What a card holds. For each email you pick, the Mailroom Clerk (Newman) writes one card: who wrote, the subject, when, a one-line gist in its own words, a suggested category, how urgent it is, and flags. The flags cover money, a password or login, a payment change, an attachment, a link, text that tries to instruct an AI, and a conversation already under way. Newman may use the mailbox's read tools and one folder where its cards land. It never drafts, sends, opens an attachment, uses the web or uses any other tool.

What the Attachment Inspector writes. For one attachment at a time, a short card: what kind of file it is, a neutral summary, facts marked untrusted, flags, and a verdict. It does nothing else.

Who reads what. In a check of your inbox, your Chief of Staff sees only the cards. The agents on one copy of the team read a cleaned copy of the one email they are working on, saved in that job's folder under a line that says it is untrusted outside text. A practice email you paste by hand in the next lesson is the one email the Chief reads itself.

Copies of the team. Each copy works on one email, in its own folder under runs/. Up to three run side by side.

Stakes levels. Routine, careful or high. High covers a complaint, anything about money, anything legal-sounding, an upset sender, or an opened attachment. Higher stakes get stronger models for the writer and the reviewer, and the reviewer is always on a model at least as strong as the writer's.

No draft, whatever the stakes. Anything about bank or payment details, passwords or logins, or someone's identity comes straight to you, marked "no draft". Nothing lifts that but your own reply.

Your own edits. Kansa's twelve checks also run on a version of a draft you edit yourself in your mail app.

Privacy and your time zone

Where customer email lands. In job folders under runs/, in the inbox record, and in the attachments folder. The ignore lines keep all of those out of save points.

The ignore lines. If some are missing, for example because you said no to them in an earlier update, Gigawatt shows you the exact lines. On your yes it adds only the missing lines at the end of .gitignore, under a short comment saying why. It never edits or removes a line you already have and never adds a line that starts with !. The same yes covers saving that one change on its own: Gigawatt saves .gitignore and nothing else.

Why the Week 2 stop happens. A file saved before the new ignore lines stays in every save point until it is untracked. Untracking only stops Git from carrying it forward. It keeps every file on your computer and deletes nothing. Copies already in older save points, or already on GitHub, stay there until you decide what to do about them, and the Chief says so plainly.

Your Langfuse records. The ignore lines keep mail out of GitHub, not out of Langfuse. If you connected Langfuse in Week 2, each job's trace keeps a copy of what the team read, so email addresses and phone numbers go there as written. If you'd rather they didn't, ask Gigawatt to mask them, ideally before you connect your mailbox: each one then becomes a placeholder such as [EMAIL-1], and a trace still shows when two emails came from the same person. Gigawatt runs a test that proves it works on your computer. Names and street addresses still go through, and your AI provider still sees what the team reads. It's your choice, and the team works the same either way.

How dates are worked out. Your profile is the file work/course/staff/owner-profile.yaml. It holds your time zone, your opening and closing times and your working days, nothing private. Gigawatt writes it on your yes and reads it back. The Chief stamps the date, time, weekday and your time zone on every check of your inbox and every job. The Support Lead (Holler) works out every "Saturday" and "next week" from that stamp and the email's own time.

How drafts are signed. The same file holds the name you sign your emails with, and, if you want them, up to four signature lines (a title, your business name, a phone number). Leave out email addresses and links: this file is kept in your save points. The writer greets each sender by the first name they signed their email with. If you would rather use their full name, or no name, tell Gigawatt and it saves that too. To change any of it later, ask your Chief of Staff to have Gigawatt update your profile.

Other email providers

The current Codex Outlook Email route is drafts only. It never uses connector sending, a browser or computer control to send. The course's historical supervised-send examples below apply only to other separately supported routes, not to Outlook. Claude Code with Outlook is not qualified here. For those other routes, this is the shape: your Chief works out how to send with the tools it has, runs the same seven checks on the draft as your mailbox shows it, shows you one line, and sends only on your word. If it can't prove a check, it sends nothing and tells you which check and what would fix it. Once a way of sending has worked for you, your Chief offers to save it as a skill of your own, so it's the same easy step every time after. Nothing is saved without your yes.

The ways it can go:

• Your email's own connection sends the saved draft. The cleanest way: what goes out is exactly the draft you approved.
• Your webmail in the app's browser. You sign in yourself, with "stay signed in" on. Your Chief opens the team's draft there, checks who it goes to, and presses Send after your yes. The browser sees whatever is on the page, so everything it reads is treated as untrusted, and it takes only the actions your Chief gives it.
• Forwarding to a separate inbox. For email no connection reaches. Your mail is forwarded into a free AgentMail inbox, where Newman reads it. Drafts and sends still happen in your own webmail, so every reply comes from your own address.

Microsoft 365. In Codex choose Outlook Email. Administrator approval may be required; the number of steps depends on your account and organization. Do not work around a refused connection by changing accounts or enabling a browser send. The team reads selected messages and creates reply drafts; no agent sends them. This lesson does not establish support for personal Microsoft or Outlook.com accounts. A personal account is not automatically forwarded to another service.

Fastmail. Its connection has a draft tool and no send tool. The team saves the draft through it, and your Chief sends it through Fastmail in the app's browser.

Other separately selected providers. An assisted forwarding route for iCloud, Yahoo, Zoho, Proton or a web host is a separate setup choice. It is not an automatic fallback for Outlook or a personal Microsoft account. If that route is the one you have separately chosen, the Chief explains the existing two steps:

• A test forward. AgentMail silently drops mail that fails its sender checks, so the Chief has you send one check email and waits for it to arrive before anything else. If it never arrives, check that forwarding is on and confirmed, and try a sender from a large provider. Some mail may never arrive through forwarding; the review page shows when the last forwarded email arrived.
• A browser profile for your webmail. You sign in to it yourself, with "stay signed in" on. The browser sees whatever is on the page, so everything it reads is treated as untrusted, and it takes only the actions the Chief gives it.

Codex Outlook Email. Your yes approves eligible drafting or a revision, never an agent send. You may send an eligible reviewed draft yourself from Outlook. Held and Practice items must stay unsent; the off switch and your written instructions cannot disable Outlook's Send button.

Who gets which mail tools

Codex Outlook Email: the exact grant plan. Gigawatt uses these tool names from the connected app; you inspect the plan rather than typing identifiers. Do not replace them with a whole connector, a wildcard, a shortened prefix or a browser route.

Sallee receives no extra Outlook grant as a side effect of this setup. Any continuing collection access is a separate exact read-only choice. On Codex, role grants are written instructions and records; do not claim they hide all other tools from the host chat.

Other mail routes where the Chief can send. Their existing roles are:

Your account's short name. The account record keeps only the short name, the provider and a fingerprint of your address. Everywhere else in the record, your account is the short name.

Your Week 2 Gmail grant. The Week 2 step could give Sallee the whole connector, send tools included, and that was the normal thing then. This week nobody keeps a whole-connector grant. On your yes, Gigawatt removes it and connects Newman read-only in its place, and the Chief says so.

Attachments. The team does not open attachments: "Attachments: Needs you, not processed." When the Attachment Inspector is turned on for your app, the lesson's table says so. Then, if your app's mail connector cannot hand an attachment to the Inspector, you drop the file into a quarantine/ folder yourself. Any attachment that names an AI product (ChatGPT, Claude, Copilot, Gemini) or speaks to an assistant comes to you under Needs you. That is the check working, not a fault.

Practice mail. The pack comes from a made-up restaurant, Hattie B's Hot Chicken (a practice example, not affiliated with Hattie B's). Every practice email comes from a reserved address that cannot receive mail, and the record marks each practice job so the send step refuses it.

Your inbox at a glance: the survey

What the survey asks for. The request is sender, arrival date and any tab or label shown, over the last 30 days, up to 300 listed items. This is not a promise of a body-free response: Codex Outlook search may include subject and message text. The Chief explains that before searching and uses only the approved scope; if you do not approve that scope, leave the survey for later. The survey makes no separate fetch, opens no attachment and does not mark, label, move or archive mail. Missing tabs or headers remain "not shown by the search". On Outlook, the items are messages, not proven distinct conversations.

Who counts. A script, not an agent. It counts listed items by sender and by domain (the part of an address after the @), by mail tab, by day of the week and by hour in your time zone, and how many look automated: no-reply and newsletter addresses and the Promotions, Social, Updates and Forums tabs. When some tabs were not shown, the automated count says "at least". Your Chief of Staff copies those numbers and never works one out.

What is saved, and where. Newman's list is counted and then deleted in the same step, and so is a list the script refuses. The counts are kept in runs/inbox/survey/, in your private records, out of your save points.

What the check refuses. Your Chief suggests two to five new or changed categories, each with its default: Team, Needs you or Skipped. It may also suggest sender domains to always skip or always bring to you. Before you see anything, a check refuses:

• a Skipped default for a category whose name or notes mention bookings, travel, orders, charges, payments, invoices, refunds, receipts, conversations, clients, scheduling or meetings, and moving any of the course's categories except vendor pitches to Skipped;
• anything but Needs you with no draft for a category about passwords, logins, verification codes, accounts, identity or bank details, however it is spelled;
• any change to the categories for security and account access, for money and payment changes, and for emails the team cannot place;
• a sender rule that names a full email address (your category file is saved in your save points, so it holds domains only), and a skip rule for a free email domain such as gmail.com, because your customers write from those too.

New categories start on the first rung of your autonomy ladder.

Your yes. You see the exact edits, numbered, and nothing changes until you say yes to them. Your Chief then writes them into work/course/staff/email-categories.md, under a "Sender rules" heading for the sender lines, and the record notes your words and the time. If you change a category's default, its count toward more freedom on the autonomy ladder starts again from that moment. You can edit the file yourself at any time; it is yours.

Run it again. Ask the same way whenever your mail changes, such as after a busy season or a new product.

Your voice package

What you can name. For example: "my sent mail to customers since March", "the About page on my website", or "the transcript of my last two podcast episodes".

How samples are gathered. Echo writes a gather request from your list. An agent that holds a mail read tool pulls the sent mail, pages are fetched by their exact addresses, and transcripts and files come from the paths you named. The samples stay in the job folder on your computer.

The three files Echo proposes. The voice package, a short brief on how agents should work with you, and the gaps. Before you approve, the Chief shows you the summary, the dos and don'ts, the relationship profiles and the gaps in plain words, and checks them for anyone else's name or number.

What is saved. work/course/staff/voice-package/ holds voice-package.md marked status: approved, empathy-brief.md, gaps.md, and approval.json with your words and the date. The folder has its own ignore line, so it never reaches a save point.

Hand edits. The saved package is fingerprinted. If you edit it by hand later, it shows as "not approved" until you say yes again.

Not your Week 1 voice. In Week 1 you may have given Echo samples for your Chief of Staff's briefing voice. That is a different thing, kept in a different file. Your email replies need your own writing.

The tools interview

The three piles. The Chief groups your answers by tool:

1. Has a connector to try now. It goes on to the three checks below.
2. Has no connector, or only one that both reads and writes. You export what Hatch needs (a weekly report, a list) into a folder you name. Hatch answers "not connected" for it and says where you would look.
3. Is a person or a paper. A fact only you would know. Hatch leaves a blank for you.

Starter sets. The course keeps a starter set for six kinds of work (coach, agency, ecommerce, restaurant, local service, course creator), three tools each, checked on 4 October 2026 against the two apps' connector directories. The Chief offers yours as a suggestion. Your answers rule, and what your app lists today wins over the course's list.

The three checks.

1. Kansa's install checklist. For each tool the Chief writes an install plan: the connector, which agent, which of its tools read and which write, and what the test pull will prove. The Quality Reviewer checks it before you are asked for a yes.
2. Read-only, by name. Gigawatt connects the agent to the read tools by their exact names, and marks every tool that creates, changes, sends, books or charges so the app asks if anything ever calls one. A connector whose read tools nobody can name is not connected.
3. One test pull. Hatch makes the smallest read that answers something from your emails, like the next three calendar events or one invoice's paid state. No write tool is called. You confirm the record is real.

The source map. One row per tool: what it answers, the exact read tools, the date of the test pull, and your yes. You say yes to the whole map at the end.

Your calendar. Read-only for the Chief and Hatch. The Chief puts a dated block from your calendar in each job, and Hatch and the reviewer read availability from it.

The summary at the end. In four lines or fewer: which tools are connected read-only, which are exports or facts only you would know, the one thing Hatch can now answer that it could not before, and that nobody can book, send or change anything through these tools this week.

Codex. Connectors belong to the chat there, not to one agent, so an agent's limit is words in its file. A connector with any write tool is held in Codex until each write tool has an approval setting. If the Chief says it is holding a tool in Codex, that is why.

Scratch and Hatch

Scratch's tools. Web search and web fetch in Claude Code, and the app's own browser for a page a fetch cannot read. In Codex, its web limits are words in its file.

Dates on every fact. Each fact carries its publication date, the date it was read, and its age at the time of the job.

What stays out of a search. No customer's name, address or order number, no amount from the email, nothing from your records. "Public events February 14 in your town", never the customer's name and party size. Every search is logged word for word.

Firecrawl or Exa, optional. Both are research connectors with free credits that let Scratch read pages and search more reliably. If you connect one in your app, Gigawatt gives it to Scratch alone, for looking things up only, through the same connect step as the tools interview. In Codex, every agent can use a tool you connect; Scratch's instructions keep web search to Scratch.

Finding the sentence. It is in the "What you never do" section of Hatch's file. The Professor can quote the section, and lets you find the sentence before it confirms.

The twelve questions

Where they come from. Whenever the Quality Reviewer (Kansa) joins a team, Kansa and Gigawatt write down what a good piece of that team's work has to be, as yes-or-no questions a reviewer can answer from the files. Your email team's review is already written: aibl-kansa-support-review.

The twelve, in order:

1. Every question in the email is answered or openly left to you.
2. Every fact traces to the evidence.
3. No invented price, date, policy or remedy.
4. No invented meeting time.
5. Every commitment fits your calendar and your policies.
6. The tone fits the sender's mood.
7. The voice matches your voice package and avoids its "never uses".
8. What the thread itself says beats any example.
9. No new recipient, link or attachment.
10. Nothing from another email or customer.
11. Every blank left for you is counted.
12. A customer can read it once and know what happens next.

Why a blank blocks "ready". A draft with any blank left for you cannot go out as it stands, so it is never READY_TO_SEND. When the blanks are the only thing in the way, Kansa says the draft needs your facts, and the Chief asks you instead of sending the writer back.

More answers when you get stuck