Work top to bottom. Each step says where it happens; click a prompt ID to open it and Copy. Her Mac Claude gets one playbook (MAC_PLAYBOOK.md) and then takes job N commands. Background and reasoning are collapsed at the bottom.
One line per step. Prompts are collapsed under the step that uses them. Her Team window is the scarce resource: Part B runs first, on Sonnet, starred chats only. Everything after Part B runs on the new account, where usage is fresh.
Where: TEAM her old account · MAC Claude Code on her Mac, new account · NEW her new account in the app · HAND by hand · OURS McLead. Start with O1. Report any step with O2.
McLead, starting the workspace setup for {{NAME}} now. Load templates/workspace-setup/SESSION_SHEET.md, PROMPTS.md and MAC_PLAYBOOK.md. The log is projects/workspace-setup/2026-10-21-first-setup/SETUP_LOG.md: record every step I report (time, result, problem) and the minutes per part. Each time I report, confirm the next step in one line and flag anything that changes the plan. Keep replies short.Step {{N}} {{done | blocked}}: {{what happened, counts, any error}}. Next?Devices: MacBook does everything. Browser tab = her Team account (Part B); Terminal and the Claude desktop app = the new account (Part C on), so no account switching all day. iPad: keep the Team account open there as a reference, switch it to the new account at handover. iPhone: new account from step 9; it's where the briefs will notify her.
Team account, Settings → Usage: note session %, weekly %, reset time. Ask which seat she has (Standard or Premium). Model: Sonnet, extended thinking off. If the window is empty, do Part C first and come back at the reset.
Terminal: paste SAVE once. From now on, Copy under a reply, then save folder/name in Terminal.
mkdir -p ~/ClaudeArchive/capture/{projects,history,settings,skills,titles}
save(){ pbpaste > ~/ClaudeArchive/capture/$1.md; echo "saved $1: $(wc -l < ~/ClaudeArchive/capture/$1.md) lines"; }
savea(){ pbpaste >> ~/ClaudeArchive/capture/$1.md; echo "appended $1"; }Copy by hand, each one → save …. Per project: the instructions field (projects/<p>-instructions), the knowledge file names (projects/<p>-files; open any file with no original on disk and copy its text to projects/<p>-file-<f>), and the chat titles from the project page with ★ on the ones that matter (titles/<p>). Non-project titles from claude.ai/recents (titles/no-project). The memory text from Settings (settings/memory-settings). One note listing skills, connectors, styles, scheduled tasks, shared artifacts (settings/notes).
New chat, no project: T1 → save settings/memory.
Two things, as complete as you can, in my original language, no preamble: 1. Everything you remember about me and my work: who I am, my company and role, people and who they are to me, clients and suppliers, preferences, how I like answers, recurring tasks, ongoing projects. Grouped bullets, small details included. 2. How I use you: the kinds of tasks (most frequent first), the format and tone I prefer, things I keep correcting, instructions I repeat, processes I walk you through step by step, and weekly or monthly tasks I could hand to an assistant.
Per project: new chat inside it, T2 with the first 15 ★ titles → save history/<p>-1. Then T2b with the next titles → save history/<p>-2, and so on. After batch 1, check Settings → Usage: that jump is the cost per batch. Size the rest to fit the window.
I'm archiving this project. For each chat title below, find it with past-chat search and write: title and date; 2-4 lines on what it was about and the outcome; decisions, with reasons; facts about my business, people, clients, prices, dates; anything still open. If you can't find one, write "NOT FOUND: <title>". Keep my original language. No preamble. End with "COVERED x of y".
Titles:
{{PASTE UP TO 15 ★ TITLES}}Next, same format:
{{PASTE THE NEXT TITLES}}★ chats outside projects: new chat, T2 with titles → save history/no-project-1.
Skills: download from Settings; else T3 → save skills/<name>. Chats whose exact text matters: copy the replies by hand → savea history/full-<title>.
Package my skill "{{SKILL NAME}}" as a zip file, exactly as it is. If you can't create files, print SKILL.md and every other file in the skill in full, each in a code block headed with its path.Done when every ★ title is covered or marked NOT FOUND, or the window is used up. Leftover batches run after the reset; nothing else waits on them.
Create the account on a separate email. Plan: Max 5× for month one, decide later from the usage diary. 2-step verification on the email. Settings → Privacy: "Help improve Claude" OFF, screenshot. Capabilities: memory on, chat search on.
Terminal: curl -fsSL https://claude.ai/install.sh | bash, new window, claude, /login (new account), /model opus. Install Google Drive for desktop and sign in. Claude desktop app and the iPhone app signed in to the new account (iPad stays on Team until step 19).
Paste M0 once. From here you only type job N: job 1 (folders, kit, Drive Company) → job 2 (inventory) → job 3 (gaps; redo anything missing from Part B).
Read {{KIT PATH, e.g. ~/Downloads/workspace-setup}}/MAC_PLAYBOOK.md and keep it as your instructions for this whole session. You are the Mac role in a Claude account migration. When I type "job N", do job N exactly as written and report in the format it asks for. Don't start any job until I ask. Confirm when you've read it, in one line.job 4 (map, 10-20 min on its own) → job 5 (decide the flagged items with her) → job 6 (load pack and quiz blocks).
Memory: Settings → Capabilities → Memory → import, paste load/MEMORY-IMPORT.txt, review the entries. No import tool? Paste it into a chat with "These are facts about me; remember them", and upload the file to a project named "About me". Projects: one per folder in load/, same name, instructions pasted, files uploaded. Skills: hers from load/skills/, then Anthropic's (skill-creator, docx, pptx, xlsx, pdf, frontend-design, canvas-design, theme-factory, web-artifacts-builder, doc-coauthoring, internal-comms), then ours.
In each project: N1.
Without guessing: summarise this project's instructions in 3 bullets, list the files you can see, and give the 3 most important facts in this project's knowledge.
In each project, N2 with its block from load/ → Copy the answers → MAC job 7 (grade) → job 8 (fixes) → re-ask the failed ones. At 20/20 she does one real weekly task start to finish. The migration is done.
Answer each question from this project's knowledge and your memory. If you don't know, write "not known". Don't guess. One line per answer, same numbering.
{{PASTE THE QUIZ BLOCK FROM load/}}Settings → Connectors: Gmail, Google Calendar, Google Drive, monday.com. Then N3. WhatsApp: her call (background agents need Evolution later); default not connected, Claude drafts her messages.
Test each connector without sending, deleting or changing anything. Gmail: find my last email from {{A KNOWN CONTACT}} and create a one-line reply as a DRAFT. Calendar: list tomorrow's events. Drive: find the folder "Company" and list its files. monday.com: list my workspaces and boards and the 3 most recently updated items. Report each as OK or FAILED with the reason.job 9 (finish the company manual with her). NEW · N4 (three Monday boards; by hand if the connector can't). One project per department, "<Dept> agent", synced to Drive Company/, instructions N5. Chief of Staff project: N6; receiving department: N6b.
Using monday.com, create three new boards. Ask me which workspace first, and don't touch any existing board.
1. "Agent Desk": From (dropdown: {{DEPARTMENTS}}), To (same), Status (New, Working, Done, Blocked, Needs founder), Due (date). Handoff text goes in each item's update.
2. "Company Ledger": Owner (dropdown: {{DEPARTMENTS}}, {{HER NAME}}), Status (open, waiting on {{HER NAME}}, waiting external, blocked, done), Priority (Urgent, Normal), Due (date), Next action (text).
3. "Job Runs": Job (text), Ran at (date and time), Result (OK, Nothing needed, FAILED), Note (text).
Show me what you created, with links.You are the {{DEPARTMENT}} agent of {{COMPANY}}. Before any task, read ORG.md, RULES.md and departments/{{department}}/AGENT.md from this project's files, then check the monday.com board "Agent Desk" for items with To = {{DEPARTMENT}} and Status = New. ORG.md's rules override everything: nothing is sent outside the company, bought or deleted without {{HER NAME}}'s explicit yes; draft instead. When you finish something, update its board item and say exactly what you changed; "done" means you saw it work. When {{HER NAME}} corrects you, propose a one-line rule for RULES.md → Lessons. Answer decision first, then only what's needed to act.{{A REAL REQUEST THAT NEEDS TWO DEPARTMENTS, e.g. "A new client just signed. Make sure they get invoiced and receive a welcome email."}} Route it per ORG.md: create the handoffs on the Agent Desk, add a Company Ledger line, and tell me exactly what you created.Check the Agent Desk for items with To = {{DEPARTMENT}} and Status = New. Take the oldest, set it to Working, do it (drafts only), then set it to Done or Needs founder with a one-line result.Scheduled tasks S1, S2, S3, S4, S5 (times are in each). Run each once now and read the output. Check the result shows up on her iPhone and that notifications are on; if not, the morning brief also goes out as a Gmail draft to herself.
Morning brief for {{HER NAME}}. Use the briefs skill, following ORG.md and RULES.md in the Drive folder Company. Sources: today's Google Calendar, Gmail since yesterday 18:00, the monday.com boards Company Ledger, Agent Desk and Job Runs. Change nothing except one row in Job Runs (Job: Morning brief; Result: OK or FAILED with the reason). If a source can't be reached, say so first.Inbox triage. Use the inbox-triage and brand-voice skills, following ORG.md and RULES.md in the Drive folder Company. Read Gmail since the last run (the last 2 hours if unsure). Drafts only: never send, archive or delete. Real actions become Agent Desk items for the right department. Add one Job Runs row (Job: Inbox triage; Result: OK, Nothing needed or FAILED; Note: counts). If nothing needed attention, say "checked N, nothing needed".
Meeting prep. Use the meeting-to-actions skill in prep mode for each of today's meetings with people outside the company: Google Calendar, Gmail history, PEOPLE.md in the Drive folder Company, and the Company Ledger. Change nothing except one Job Runs row (Job: Meeting prep).
Weekly review for {{HER NAME}}. Use the briefs skill: Company Ledger, Agent Desk, Job Runs and Gmail from the past 7 days. Change nothing except one Job Runs row (Job: Weekly review).Watchdog. In the monday.com board Job Runs, check today's rows. Expected: Morning brief 1, Meeting prep 1, Inbox triage {{N}}, Weekly review 1 on Sundays. For each missing or FAILED job, create a Company Ledger item (Owner: Chief of Staff; Priority: Urgent; Next action: check why <job> didn't run). Add a Job Runs row (Job: Watchdog). If everything ran, say "all jobs ran" with the counts.N7 (brand voice from sent mail) → MAC job 10 → upload the zip. N8 (first skill). N9 (first SOP).
Use Gmail to read my last 40 sent emails (skip one-word replies and forwards). Fill in the attached brand-voice skill template from them: how I write, real phrases I use, what I never say, typical length, greeting and sign-off, and 3-5 short real examples, one per kind of email. Then ask me for the brand colors, fonts and where the logo is. Give me the finished SKILL.md as a downloadable file.
Use the skill-creator skill. Build a skill for this recurring task of mine: {{PASTE THE TOP ITEM FROM map/SKILL-CANDIDATES.md}}. Interview me briefly, write it, then test it on a real example of mine.Use the sop-builder skill. I'll explain how I do {{PROCESS}}. Turn it into an SOP, save it in Drive Company/SOPs, and add it to SOPs/INDEX.md.job 11 (handover note, includes the usage diary). HAND · she pauses and resumes one task, disconnects and reconnects one connector. Sign the iPad into the new account.
O3.
The setup is finished. Notes: {{what surprised you, what took longest, what was skipped}}. Fill the minutes per part in SETUP_LOG.md, fold every lesson into templates/workspace-setup/ (sheet, prompts, playbook, checklist, README), and list what to fix in the kit before the next setup.Total about 5 to 5.5 hours with lunch. Short on time? 17 to 19 need only her new account and can be finished remotely.
/usage). After two weeks: session limit hit twice a week or weekly at 70% → up a tier; weekly under 35% two weeks running → down a tier.save line in Terminal. With many chats, start with the projects that matter most and let the rest run while Phase 3 starts on what's already captured.How OctoMonic runs its own AI team, checked 20 Oct 2026: 7 agent personas, 22 skills, 59 rules learned from corrections, 62 decision files and about 140 scripts. Here is what carries over to a company run by departments.
An audit of how OctoMonic runs its own AI team (AIOS, checked 2026-10-20: 7 agent personas, 22 skills, 59 feedback rules, 62 decision files, ~140 scripts), turned into what a client setup should copy, adapt, or skip.
| OctoMonic (AIOS) | A founder with departments | |
|---|---|---|
| Unit of work | A project (Monesys, GoMobile, Beertza) | A department (Sales, Ops, Finance, Marketing) |
| Agents | One lead (McLead) + one Claude per project, delegating to subagents per task | One standing agent per department with narrow, repeatable jobs, plus a Chief of Staff |
| Where agents live | Code workspaces (Claude Code) + one always-on VM agent (Operator) | Claude Projects (interactive) + cloud scheduled tasks (autonomous) |
| How agents talk | Context packets + session log tags ([for: X]), read at boot | The Agent Desk board in Monday, read at the start of every task |
| Live state | Git files (ledger, log, board YAML) | Monday boards (cloud jobs can write there; they can't write to her Mac) |
| Documents | Git repo | A Google Drive "Company" folder, synced into Projects |
The pattern carries over. The plumbing changes: her agents don't run code workspaces, so git becomes Drive (documents) and Monday (state).
| AIOS practice | Why it works | In her setup |
|---|---|---|
One master file (CLAUDE.md) read by every Claude, then a persona file per agent, then project files | Every agent starts from the same facts; a change lands everywhere at once | ORG.md → departments/<x>/AGENT.md → project knowledge |
| Two state surfaces only: the ledger (open) and the log (history). We collapsed from 9 | Few places to look; nothing open hides in a third file | Monday "Company Ledger" + the log in Monday updates / LOG.md |
| Handoffs are written, with To/From, never "remember to tell X" | The receiver acts without asking; nothing lives in one chat's memory | Agent Desk items with From/To/Ask/Why/Context/Done-when |
| Hard gates: nothing external sent, nothing spent, nothing deleted without an explicit yes | Lets autonomy run without fear; every mistake is a draft, not an incident | ORG.md rules 1-3, repeated in every scheduled task prompt |
Corrections become rules (memory/feedback_*.md with Why / How to apply) | The same mistake doesn't happen twice, across every agent | RULES.md → "Lessons": an agent corrected proposes a one-line rule |
| Decisions get a file with the reason | Nobody re-argues a settled question; agents check before proposing | decisions/YYYY-MM-DD-topic.md in Drive |
| Briefs: morning, end of day, weekly | One place to see everything; decisions surface instead of hiding | Morning brief + weekly review jobs |
| Silence is not success (monitor.py; "checks that return empty success"; "silent failure outlives redundancy") | Our worst outages were jobs that quietly stopped (OpenAI + Gemini both dead, nothing alerted) | Every job writes a line to a "Job Runs" board; a watchdog job flags any job that didn't run |
| Run-and-show-output / persistence proof: "done" needs evidence | Kills the agent that says "saved" when nothing was saved | ORG.md rule 4; the company-brain skill names the file it wrote |
Skills registry (SKILLS.md) as single source of truth | Agents know what tools exist; nothing is rebuilt twice | Company/SKILLS.md |
Connections registry (connections.md): every tool, status, how to reach it, never the secret | No agent re-discovers an integration; no secret leaks into a file | Company/CONNECTIONS.md: what each app may and may never be used for |
Voice profile from real messages (refresh_voice_profile.py) | Drafts sound like Tom, not like an AI | brand-voice skill filled from her real sent mail |
| Natural language, not commands | Tom never memorises syntax | Skill descriptions trigger on how she actually talks |
| Deliverables as shareable pages (mockups) | Reading experience matters; links forward easily | mockup-page skill |
Weekly self-test (self_test.py, /team-skill-audit, /level-up) | Drift gets caught before it bites | A weekly "system check" prompt in the Chief of Staff (step 10.4) |
| Monday write safety: write only to items you own | A shared workspace token can damage colleagues' work | CONNECTIONS.md: agents write only to Agent Desk, Ledger, Job Runs, and items assigned to her |
| TLDR first, local time | How Tom reads; zero mental conversion | "How she likes answers" in ORG.md, from her captured working style |
| AIOS practice | Her version |
|---|---|
session_start.py boot routine (pull, inbox drain, brief) | The first lines of every department's project instructions: read ORG.md, RULES.md, own AGENT.md, then its Agent Desk items |
| Context packets drained at boot | Agent Desk items with status New, picked up at the start of each task and by the hourly job |
structure.yaml board with tiers, ranks, one expedite slot | Monday Ledger with a Priority column; at most one "Urgent" at a time (our expedite cap) |
| Operator, always-on on a VM, answering WhatsApp | Cloud scheduled tasks (no server). WhatsApp only if she chooses a bridge (Evolution, as Operator uses) |
| Subagents per task | Inside a department, Claude Code / Cowork can still fan out; she never needs to think about it |
state_sweep.py detecting done-but-not-recorded | The watchdog + weekly system check: Ledger items with no update in 7 days get chased |
RULES.md short; merge lessons monthly in the system check.Everything from the Team account, proven by the quiz. Training off, memory imported, on desktop, phone and Claude Code.
Gmail, Calendar, Drive and monday.com, each tested with a real read and a draft. WhatsApp only if she chooses a route.
Her existing skills unchanged. Anthropic's design, document and skill-creator skills. Nine OctoMonic starter skills, and her own first skill built on the day.
ORG.md, RULES.md, an AGENT.md per department, SOPs, SKILLS.md, CONNECTIONS.md and decisions. Each department Project syncs these files.
Company Ledger (everything open), Agent Desk (every handoff between departments), and Job Runs (proof each job ran). Cloud jobs can write to all three.
Morning brief, hourly inbox triage, meeting prep, weekly review, and a watchdog that flags any job that didn't run. Real-time triggers come later.
Checked 20 Oct 2026. The first four rows are from Anthropic's own help center, updated this month. The rest are from secondary sources and get verified live.
| Fact | What it means for tomorrow | Confidence |
|---|---|---|
| Anthropic moves personal accounts into Team, one-way. There is no path back out. | Nothing migrates automatically. We copy, re-create and prove parity. | Official |
| A personal account and an organization account can share one email. You switch from the menu bottom left. | She keeps full access to the Team account during and after the move. That is how "access to everything she had" is guaranteed. | Official |
| Member export needs an Owner to turn it on. It covers her own chats, not the contents of uploaded files. A Primary Owner can export the whole organization. | Must happen tonight. Project knowledge files are re-uploaded from the originals on her disk or Drive. | Official |
| When accounts move into Team, chats, projects, files, memory, Drive and repo sync settings and Claude Code settings all carry over. | That list is the inventory checklist for going the other way. | Official |
| Free, Pro and Max run on consumer terms, where the model-training setting defaults to on. Team runs on commercial terms with no training. | Switch "Help improve Claude" off first and screenshot it. Settle what stays in the Team account. | Verify live |
| Claude can import memory: you paste the memory text into the import tool, then review the entries. | Copy her Team memory and paste it in. Check every entry. | Verify live |
| Cowork scheduled tasks run in the cloud for Max, so the machine can be off. | The morning brief, hourly triage and meeting prep need no server. | Verify live |
| Claude Code routines run in the cloud on a schedule (hourly at most), on an HTTP API call, or on GitHub events. | Real-time triggers are possible: Monday or Gmail fires a small relay, and the relay calls the routine. | Verify live |
| There is no official WhatsApp connector. The options are a WhatsApp Business API provider (official, needs a business number and Meta setup) or unofficial bridges that log in as her own WhatsApp (an afternoon to set up, but the number risks a ban). | WhatsApp is not connected on day one by default. Claude drafts her WhatsApp messages, and she decides between the two routes knowing the trade-off. | Secondary |
| The connector directory includes Gmail, Calendar, Drive, monday.com and Slack. What each one can write varies. | Test one real read and one real draft per connector before relying on it. | Verify live |
| Route | What it is | Good for | Catch |
|---|---|---|---|
| Official Business provider | WhatsApp Business Platform via Meta or a provider like Twilio | Customer-facing messaging | Usually a separate business number. Messages she starts outside a 24-hour window need approved templates |
| Evolution API on a server | What Operator runs on: a server linked to her WhatsApp, with webhooks | Agents reacting to messages in the background | Unofficial, so the number risks a ban. Needs a server. Half a day or more |
Local bridge (whatsapp-mcp) | Runs on her Mac and plugs into Claude | Asking Claude about her chats while she's at the Mac | Unofficial, so the same ban risk. Only works while the Mac is on |
Default for the day: not connected, and Claude drafts her WhatsApp messages. If she wants agents in the background, Evolution is the follow-up, using a number she can afford to lose.
The eye-openers. Day goes live tomorrow, Show is demonstrated on her real data, and Next is quoted as a follow-on.
| What | What she sees | |
|---|---|---|
| Morning brief | One message before her day starts: meetings with prep notes, what's waiting on her, what ran overnight and what failed. | Day |
| Inbox that triages itself | Replies already drafted in her voice. Real actions become board items, and nothing goes unanswered for more than 3 days. | Day |
| Meeting prep packs | Before each meeting: who they are, the last emails, open items, what to ask. | Day |
| Agents talking to each other | Sales closes a deal, Finance gets a handoff to invoice, and she watches it happen on one board. | Day |
| Ask the whole company | "What did we promise this client in March?" answered from her Drive and every past chat. | Day |
| Weekly CEO review | Sunday evening: what moved, what's stuck, three decisions to make. | Day |
| Meeting to action items | Notes or a transcript become Monday items with owners, plus a follow-up email draft. | Show |
| Voice to task | A voice note from her phone becomes a task on the right department's list. | Show |
| Documents from templates | Proposals, quotes and agreements built from her templates as a reusable skill. | Show |
| Real-time agents | A Monday change or a new email wakes the right agent within a minute instead of within the hour. | Next |
| Finance capture | Invoices and receipts from email filed and summarised. Read-only toward her accounting system. | Next |
| Watchers | Competitors, tenders and price changes checked on a schedule, with only changes reported. | Next |
Our own hard rules, carried over. They are written into her ORG.md, and they're why autonomy is safe to switch on in one day.
| If | Then |
|---|---|
| Not using the owner-enabled export (the case here: reaching the owner is a formality we're skipping) | Claude writes up its own history inside the old account, driven by the chat titles. Summaries, not transcripts. Chats that need their full text are copied by hand. No scraping around the export: Anthropic's terms ban automated access to claude.ai. |
| Past-chat search is off or restricted by the organization | Capture by hand, project by project, starting with the projects she uses most. |
| Her Google Workspace blocks Claude | The admin allowlists the app (2 minutes). If the admin isn't available, connect Calendar and Drive first and add Gmail after. |
| Cloud scheduled tasks behave differently than reported | Use Claude Code cloud routines on a schedule. As a last resort, desktop schedules that need the machine awake. |
| She hits the usage limit in week one | Move to 20×, or make triage run every two hours. |
| Business data on consumer terms | Training is off. Material under NDA or owned by the organization stays in Team. All of this is written into the handover note. |
| Claude Design opens in the wrong account (reported bug when both accounts share an email) | Check the blue checkmark before working. If it keeps happening, use a separate email. |