NEEDS: the router file (Memo 01) · a handful of messages you’ve already sent · projects (Spec Sheet 05)

Your AI already writes a clean reply. A voice file makes it sound like you. Asked for my answer to Ambervale’s* board treasurer on the reserve question, a plain chat opened with her name and ran 63 words. In a project holding my voice file, built from eight messages I’d already sent, the same request came back in two sentences and 30 words, the way I write to the board.

GRAY ZONE

A voice file gets the first draft close to how you write, and the draft still comes back to you for the read.

Eight sent messages become a voice file, the file goes into a project, and one request comes back two ways.

The setup

The file is built from messages you’ve already sent. Each rule is one line, with the number of the message behind it.

REGISTER BY READER

How you write to each person or group. Where a reader has one message or none, it says the register is thin.

SENTENCE RULES

Length, openings, and how numbers and dates appear.

NEVER

What you’d never send. A ban goes in the day a draft comes back with the habit.

ALWAYS

What every message carries, such as the figure that matters, up front.

SIGN-OFFS

How you close, by reader, including not at all.

TWO EXAMPLES OF ME

Two messages, word for word.

ADDED SINCE

Rules added after the build, dated, newest last.

Read it before you use it. Ambervale’s file made a rule out of one question I’d ended with a period, so that line goes. It had one message to a board member, so I add the rule it couldn’t see: full sentences for the board, dated.

I once kept a ban on em dashes in a file of standing rules, and a draft came back full of them. The file wasn’t loading in that session. That was a setup gap, and it’s why the voice file now lives where it loads every time. A project is a workspace whose files and instructions every chat inside it starts with. Put the voice file and the router file in one, with an instruction to follow the voice file; Spec Sheet 05 has the steps. Scheduled runs keep reading the shared folder from Memo 04.

No projects on your plan or work account? Attach the voice file to each chat, the way Memos 02 and 03 attached the router file. It costs a step every time, and each chat reads the copy you attached.

HUMAN LOCK

Both project drafts stopped at the same place: what to tell Priyanka about December. The router file says the reserve frame isn’t the draft’s to pick. Rennick settled it at Thursday’s check-in, so I write that sentence myself: a one-line report on Dec 17, with Warren’s four quarters under policy in it. The file gets the draft close, and what it says stays mine.

The close

The rewrite I used to do on every draft, taking out the greeting and moving the number up, is done before the draft arrives. The time goes to the one sentence only I can write. That’s your AI building what you describe: you name the decision and who it’s for, and the draft starts where your edit used to end.

What’s one phrase your AI keeps putting in your drafts that you’d never send?

* Ambervale is a fictional nonprofit I use in these memos so I can show real workflows without showing anyone’s real inbox.

Ops Bureau · Memo 05 · 2026-10-29

The rule to keep is on page 3. On ChatGPT, every member of a shared project sees its chats and files, and deleting a file removes it for everyone. On Claude, sharing is a Team and Enterprise feature, with view or edit access. A shared Drive project in Google shares only the list of sources, and each person’s conversations stay private. Under the table sits the human lock: decide what goes in before you share.

Page 4 is the routing. Scheduled runs keep reading the shared folder they already use: a Claude Cowork project can hold its own scheduled tasks, though the pages don’t say they read its files, and a ChatGPT task made in a project can’t open the project’s files. A board-prep project thinks with you. A grant-report project with your voice file builds what you describe. Add the router file and the voice file from the shared folder. Claude keeps Google Docs from Drive synced in private projects; anywhere else, replace the file when it changes.

Sources, all checked Oct 26, 2026. Anthropic, Claude Help Center: What are projects?, How can I create and manage projects?, Organize your tasks with projects in Claude Cowork, Schedule recurring tasks in Claude Cowork, Upload files to Claude, Use Google Workspace connectors. OpenAI, Help Center: Projects in ChatGPT, Scheduled tasks in ChatGPT. Google: Organize your projects with notebooks in Gemini Apps, Use Gems in Gemini Apps, About the transition from Gems to skills, Organize, save, and share your Gemini sources using projects in Drive. Microsoft Support: Add references to your Microsoft Copilot Notebook, Provide custom instructions for your Microsoft Copilot Notebook.

Checked against the sources Oct 26, 2026 and again on Oct 29, 2026. Claude’s pages describe the current version of projects while a new one is in beta. Limits and plan names are as printed on the day. If a row on this sheet stops matching its source, this post gets a RETIRED line at the top.

NEEDS: the router file (Memo 01) · the morning stack (Memo 02) · the decision queue, with shared documents connected (Memo 03) · scheduled runs (Spec Sheet 04)

A check-in should open on the calls the other person has to make. On Ambervale’s* desk this Thursday there are five, and the draft that lists them is built at 8:30 by a run that nobody starts. Before then I read the morning brief and overrule the queue on the lease, and the check-in that comes back puts the five calls on top, in the executive director’s order.

RUNS WITHOUT YOU

A check-in is assembled from a week of messages your AI already reads, so it can be built on a timer, and the draft waits for you.

Three runs on a schedule, a check-in built at 8:30, one pushback line rewritten before it goes anywhere.

The setup

The sections always run in this order.

DECISIONS I NEED FROM YOU

Every call on their list, in their order, each with your recommendation, a needed-by date, and the pushback you expect, with your answer.

DECISIONS I MADE

Your own calls from the week, each with the date it can still be reversed.

MOVED

Deliverables and deadlines, three lines at most.

WAITING

Who, what and since when, leaving out anything above.

WATCHING

One risk that isn’t a decision yet.

The first line names the meeting and its date, so you can check the week. If the first section is empty, it offers to cancel.

A scheduled run has nobody there to attach files, so the router file and saved files need a home it can read: here, a shared folder, through Memo 03’s documents connection. Save them as Google Docs or Word files. Each scheduled task is one line pointing at its file: the brief at 7:00, the check-in at 8:30. Each run is its own session, so the queue’s override reaches the check-in through the dated line in the router file. A run that stops to ask permission waits for an answer, so before the first run, allow its reads and set anything that writes to ask first.

A scheduled brief of mine once made the same formatting bug four mornings in a row, and I fixed the output by hand each time. That was a setup gap: on a schedule the fix goes in the saved file, and the next run gets checked once by hand.

HUMAN LOCK

The run’s pushback line on the instructor offer is about cost. The question I know the executive director will ask is who tells Beatriz, the lead instructor whose raise comes with the offer, and when. I strike the run’s line and write his question, with my answer, before the check-in goes to him. A scheduled run reads and drafts, and nothing it makes leaves my desk until I’ve read it.

The close

The three runs finish early, and the time before the meeting goes to the part only I can write: the answers to the pushback, and whether each recommendation is one I’d defend in the first five minutes. That’s the desk running without you, and the meeting it feeds opens on decisions.

Which recurring meeting on your calendar opens with updates, and what decision is sitting at the bottom of it?

* Ambervale is a fictional nonprofit I use in these memos so I can show real workflows without showing anyone’s real inbox.

Ops Bureau · Memo 04 · 2026-10-22

The rule to keep is on page 3. ChatGPT's page says a task that needs approval pauses until you review it. Google's says Gemini is designed to ask before it sends, changes your data, purchases, submits a form, or signs in with Google, and that a schedule running while you're offline may complete an action you can't stop. Claude's Cowork pages describe three modes and an approval mode field on the scheduled task, and say nothing about a run that pauses with nobody there; the Claude Code Desktop page says a local task in Manual mode "stalls until you approve it." Grok's and Copilot's pages say nothing about approvals. Under the table sits the human lock: a scheduled run reads and drafts, and nothing it makes leaves your desk until you've read it.

Page 4 is the routing. The morning brief and the Thursday check-in draft run without you, from a calendar and inbox the run can only read. A Friday run that lists the week's open questions with their dates thinks with you. A monthly board update drafted from the program tracker is built from what you describe. Every run reads, then saves a draft; before the first run, allow its reads and set anything that writes to ask first.

Sources, all checked Oct 21, 2026. Anthropic, Claude Help Center: Schedule recurring tasks in Claude Cowork, Use Claude Cowork safely, Get started with Claude Cowork, Use Claude Cowork on Team and Enterprise plans; Claude Code docs: Automate work with scheduled tasks, Schedule recurring tasks in Claude Code Desktop. OpenAI, Help Center: Scheduled tasks in ChatGPT. Google, Gemini Apps Help: Schedules in Gemini Apps, Create & manage schedules for tasks in Gemini, Manage your tasks & workflows in Gemini Apps. Automations in Grok, x.ai. Microsoft Support: Schedule your most used Copilot prompts.

Checked against the sources Oct 21, 2026. Two Google pages disagree on whether a task schedule runs with the device off; both are cited. Caps and plan names are as printed on the day. If a row on this sheet stops matching its source, this post gets a RETIRED line at the top.

NEEDS: the router file (Memo 01) · the morning stack (Memo 02) · your chat and shared documents, connected (Spec Sheet 02) or pasted in and attached

Memo 02 ended on the item waiting on you. On Ambervale’s* desk, it’s a message the executive director sent at 5:40 Wednesday evening, with five calls for Thursday and one instruction: “Come with positions, not summaries.”

Those five calls are buried in 51 messages across six channels and three direct threads, four documents and two forms. A decision queue reads it all and returns each call with a recommended position.

HUMAN LOCK

Your AI can sort and price every call in the week, and each one is still made by the person who answers for it.

51 messages read, five calls returned, one recommendation overruled.

The setup

The queue holds the calls you make and the ones you owe a recommendation on.

OPTIONS

Two to four real choices, including doing nothing when waiting is one of them.

RECOMMENDED

One pick and its strongest reason. It’s the position you take into the room, or strike.

COST IF WRONG

Hours, dollars or a relationship, and whether it can be undone. It sets the order, so the call you can’t take back comes first.

NOT MINE

Calls that belong to someone else, with the owner from your router file’s WHO DECIDES WHAT.

NOT DECISIONS

Updates and plain tasks, one line each, so you can see what it set aside.

Save these fields in a file next to your router file, and attach both when you start the queue.

The queue reads the week where it happened: your chat channels and shared documents, through the same one-time sign-in as your calendar and inbox. Some of those connections can post or edit as well as read. The queue only reads, so set anything that writes to ask first. For now you start it yourself; scheduled runs come in Memo 04.

The queue also works without any connection. If you aren’t ready to connect, aren’t sure how yet (check out Spec Sheet 02), or your work account doesn’t allow it, paste the week’s messages into the chat and attach the documents and forms as files. The queue reads pasted text and attached files the same way. What changes is the gathering: you collect the week yourself, and each run needs a fresh copy.

THE OVERRIDE

The top call in the video is a five-year lease the board must approve, and the landlord wants an answer four weeks before the board meets. The queue recommends a special vote. I strike it and ask the landlord to hold, knowing he may walk. The reason is one line no file holds: Ambervale’s chair takes multi-year commitments only at regular meetings. It goes into the router file, dated.

The close

The queue spares you the second pass through the week, the one spent hunting for what needs deciding. On Ambervale’s desk, that attention goes to the lease, the one call that turned on knowing the chair. That’s the queue thinking with you: it put the hold on the list of options and gave you a recommendation worth arguing with.

What’s the oldest call on your desk right now, and what would it cost to be wrong about it?

* Ambervale is a fictional nonprofit I use in these memos so I can show real workflows without showing anyone’s real inbox.

The rule to keep is on page 3: read the frame before the numbers. Who ran the test, which models are missing, what setup was used, how big the gap is, what a blank means, which version it is, and which number is shown. Under it sits the human lock: which model your desk trusts is your call, made on your own task.

Page 4 is the routing. The gift-to-thank-you workflow across your apps runs without you, with every write set to ask first. Four years of board minutes, held at once to trace a policy change, thinks with you. A board update from a quarter of program data, or a sign-up tracker from a paragraph, is built from what you describe.

Sources, all checked Oct 14, 2026. Launch tables, for the lines that say where each test appears: Anthropic, Introducing Claude Opus 5.5; Google DeepMind, Gemini 4 Argon model page and evaluation methodology; OpenAI, GPT-6 Astra; xAI, Introducing Grok 4.7. Builders, for what each test measures: Zapier, AutomationBench; XLANG Lab, OSWorld 2.0; OpenAI, GDPval, GraphWalks and MRCR; Artificial Analysis, GDPval-AA; UC Berkeley RDI with Snorkel AI, Agents’ Last Exam; Vals AI, Finance Agent, Vibe Code Bench and Vals Index; Harvey, Legal Agent Benchmark; Surge AI, Chartography. The token rule of thumb is OpenAI’s: Understanding and counting tokens.

NEEDS: the router file (Memo 01) · your calendar and inbox connected, or pasted as text (Spec Sheet 02)

8:15 at Ambervale’s* desk. 8 tabs open: the calendar, the inbox, 2 chat channels and 4 shared docs.

Each one holds a piece of what today needs from you. Stitching them together eats about 40 minutes before you make a single real decision.

A morning stack does that assembly with one command. You run it when you sit down, the brief is on the page a few minutes later, and the 40 minutes go into strategizing.

GRAY ZONE

A morning brief is a real draft of your day. It’s the first thing you run, before anyone acts on anything in it.

Eight tabs closed, one brief built, three lines edited.

The setup

The morning command is short: 5 parts with one line each.

DATE

goes in the one-line request you type each morning, so “today” and “tomorrow” mean the same thing to you and your AI.

SOURCES

The calendar, inbox, channels and folders it reads, by name. Anything you don’t name, it isn’t asked to read.

NEVER READ

A pointer to the STANDING RULES heading in your router file, so the never-read list lives in one place.

SECTIONS

The brief comes back in 4 parts: today’s calls, what moved since yesterday, what’s waiting on you, and one flagged risk. Any date the brief infers gets a [?].

WHERE IT LANDS

As text in the window you already have open. If you have to open a file to read the brief, you’ll read it later (or never).

Your AI reads only the accounts you connect, through a one-time sign-in that Claude calls a connector and ChatGPT calls an app. Who has to approve it, and how much it may do without asking, is on Spec Sheet 02. If you can't connect, copy your calendar's day view and your inbox list and paste them under the command. The brief works the same way on pasted text.

HUMAN LOCK

In the video I delete 2 lines and rewrite a third before the brief goes anywhere.

Ambervale’s Thursday brief shows a funder report due Friday the 14th. The brief flags it with a [?], because November 14 is a Saturday.

The brief can catch the mismatch. Which day the report goes out, and whether the funder hears about the change, is your call.

The close

For now you run the brief yourself: one command, first thing. A later memo puts it on a schedule, so it’s waiting when you sit down.

The 40 minutes it hands back go to the item that’s waiting on you, the one that needs options and a recommendation before anyone can decide it.

That’s where the next memo starts: the decision queue. It puts the options, your recommendation and the cost of being wrong on one page, so the call gets made this week.

What’s the first tab you open every morning, and what are you looking for when you open it?

* Ambervale is a fictional nonprofit I use in these memos so I can show real workflows without showing anyone’s real inbox.

The rule to keep is on page 3: what your AI may read is your call, and on a work account it is also IT’s, so it gets decided before the first sign-in.

Connected with read access only, a calendar and an inbox give you a morning brief that runs without you, a question that thinks with you (“what did I commit to by email that isn’t on my calendar?”), and a table built from what you describe: every funder report date in the inbox, with its thread.

The sign-in steps, in Claude and in ChatGPT, are on the vendor pages listed under Sources.

Sources: OpenAI Help Center, “Connected apps in ChatGPT” and “Connecting and managing app accounts in ChatGPT.” Anthropic, Claude documentation, “Get started with connectors,” “Google Calendar” and “Gmail.”

First posted on the Ops Bureau Company Page Oct 6, 2026. Quotes checked against the sources Oct 5 and again Oct 8, 2026: unchanged.

Ops Bureau is independent and not affiliated with OpenAI or Anthropic.

Start on Sonnet for most work. Drop to Haiku for simple, repetitive tasks, and move to Opus for complex work you'll question and redirect as you go. Move up when a model struggles, and move down when a cheaper one holds up. Then check your own desk for the habit of sending every task to the same model: the inbox sort can go to Haiku, the budget scenarios to Opus, and the grant report to Fable.

Source: Anthropic, “Choosing the right Claude model: Haiku, Sonnet, Opus, or Fable.” Claude Academy. academy.claude.com/tutorials/choosing-the-right-claude-model

First posted on the Ops Bureau Company Page Sep 29, 2026. Quotes checked against the source Sep 28 and again Oct 7, 2026: unchanged. Model lineups change. If this guide changes, this sheet gets marked retired here.

Ops Bureau is independent and not affiliated with Anthropic.

I asked for the next board meeting’s agenda twice, with the same eight words. The first answer came back as a blank template: bracketed blanks for the organization name, stock consent and discussion block. No indication that it knew Ambervale*. The second put December 17 on the page, pulled names from the roster, and flagged that the file’s next full-board date and a strict reading of “next month” do not line up. The only difference between the two runs was a file I attached the second time.

The same request, without the router file and then with it.

ASSEMBLY

An agenda is assembly work. Every fact it needs should already exist somewhere in the organization. Once those facts sit in one file, the draft does not ask you to dig them up again.

The setup

A router file is one document, linked to every session, that answers what a new hire would ask in week one. To have your AI load it without being asked, attach it to a Project in Claude or ChatGPT, so every conversation in that Project has it. With any other AI, paste it at the top of the session. Eight headings.

ORG

What you do, for whom, at what size. One paragraph, in numbers where you have them.

FISCAL YEAR

The dates, so “next quarter” means the same thing to both of you.

PROGRAMS AND NUMBERS

Each program, its share of the budget, the one metric you watch.

BOARD AND COMMITTEES

Who sits where, when each meets, and what each is allowed to decide.

RECURRING DELIVERABLES

Everything with a due date that repeats, funder reports included, with the funder named.

WHO DECIDES WHAT

Decision rights, by role. Where nobody has decided, the file says so.

READERS

One clause per audience on who they are to you. For Ambervale, the board is nine volunteers who read on the train.

STANDING RULES

Where output lands, what has failed before, what your AI may never read, and the rules learned the hard way, dated.

The test for a finished router: attach it, ask for something routine, and see whether the answer comes back with a question you already answered in the file. If it does, that heading is thin.

HUMAN LOCK

WHO DECIDES WHAT decides what an agenda can hold. The file knows Ambervale’s operating reserve sits at 2 months against a 3-month policy, and it will put the reserve on the December agenda. Whether the board sees it as a decision or as a report is yours to call, and the file waits until you’ve made it.

The close

The minutes that used to go to re-explaining my own organization at the top of every session now go to the one item on that agenda that needs an argument. On Ambervale’s desk that’s the reserve conversation, which the file can put on the agenda and can’t settle. The draft runs without you. The hour it hands back gets spent on a call.

Once this file attaches every time, the next leverage is what the machine can read without you attaching anything: email, calendar, Drive.

What’s the one fact about your org you’ve typed into a chat window more than 3 times this month?

‌* Ambervale is a fictional nonprofit I use in these memos so I can show real workflows without showing anyone’s real board.