What you'll have at the end

Three things, every day, without you doing them.

1. A morning Telegram message at 8am. "Three external meetings today. Briefs ready. Highest priority: [name] at 10am. Power 3 tasks: [...]." You read it on your phone. You're prepped before coffee.

2. A recap email staged in your drafts within 30 minutes of every meeting ending. Written in your voice. Action items already added to your tasks. CRM already updated. You read, you hit send.

3. A "messages need eyes" Telegram nudge a few times a day. "8 messages from Tier-1 humans across email, iMessage, WhatsApp, LinkedIn. 2 drafts ready to send. 1 needs your judgment." You open one feed. You triage in 7 minutes. What used to be 90.

Who this works for. Who it doesn't.

Honest success rates from watching founders try this in 2025-2026.

If you're in the bottom three rows, this playbook will not save you. Either upgrade your environment first or hire a human EA.

If you're in the top three rows, the rest of this document is for you.

AI assistant vs human EA vs both

Before you spend a Saturday building this, decide if you should.

The right answer for most early-stage founders: build your own first. Add a human EA later when "Sasha's queue is full and I'm still the constraint." For founders post-Series A with heavy in-person calendars, both.

What you delegate. What you keep.

The single biggest reason founders quit AI assistants in week 2: they tried to delegate judgment work, got bad drafts, rewrote, and concluded "AI doesn't work."

The framework:

Apply this filter before every prompt. If it's judgment work, write it yourself. The AI is your scaffolding, not your substitute.

How long this takes

Pair sessions matter. Solo founders who try this end up burning the day on OAuth errors. Pair with a buddy who has done it before, or a coach.

Step 0. Hardware and software baseline

Get this set up before Pair Setup. Most of it is one-click.

Recorder (the most important hardware decision)

Most founders need only Granola. Don't over-buy.

Other hardware

Software baseline (install before Pair Setup)

1. Claude Cowork (Mac) or Claude Desktop (Windows). Free at claude.com/download.
2. Beeper Desktop. Today's recommendation. See "Beeper failover plan" below in case Beeper changes posture.
3. Telegram on your phone.
4. Granola if you take digital calls.

Beeper failover plan (be honest about the dependency)

Beeper bridges iMessage, WhatsApp, LinkedIn, IG, Matrix, SMS, Telegram, Discord, Signal in one inbox. It is the right answer today. It is also a YC-funded company that depends on bridges Apple has shut down before. If Beeper changes: - iMessage failover: native Mac chat.db (Mac only) or move to a different DM channel - WhatsApp failover: WhatsApp Business API - LinkedIn failover: Unipile

Plan B documented. Don't lose sleep. Use Beeper today.

What you can't connect right now

• Apple Mail. No reliable native connector. Forward to Gmail or migrate.
• X DMs. Locked down by X in 2024 to 2025. Manual only.
• Customer Slack workspaces (you're a guest). Read-only at best.
• Roam. No bridge yet. Manual feed.
• iMessage on Windows. Beeper bridges via iOS but quality is lower than Mac native. Trade-off you accept.

Order of operations (Step 0)

1. Order Plaud Note only if your meeting mix is mostly in-person/phone.
2. Install Cowork or Claude Desktop (10 min).
3. Install Granola if you take digital calls (5 min).
4. Install Beeper, log in to channels (20 min).
5. Install Telegram on phone (1 min).
6. Schedule Pair Setup.

Pair Setup session (4 to 6 hours)

You'll knock out: decisions, folder, CRM tier setup, first connectors, voice (two-pass).

Hour 1. Decisions and CRM tier

Decision #1. The name

Your assistant needs a name. The 100x agent naming guide (Appendix L) has 8 laws. Short version:

Quick picks:

Don't agonize. Decide in five minutes.

Decision #2. The sign-off name

Every outbound message ends with a name. Pick which: - Your name (recipient thinks you wrote it) - Your assistant's name (transparency) - Separate outbound voice (Nik uses "Nikolai" for outbound, "Sasha" for the OS)

Decision #3. The stack

Decision #4. Single agent or split?

• Single Sasha. One assistant runs everything. Recommended for solo founders.
• Sasha + Nikolai. OS layer is "Sasha" (private). Outbound voice is "Nikolai" (public). Recommended for founders with heavy outbound cadence.

Decision #5. Working hours and channels

Decision #6. CRM tier setup (the most important step)

This is the load-bearing decision. The single failure mode that has burned every founder we've watched is "Tier-1 person messaged on a quiet day, filter scored on volume, message went unanswered for 4 days." The CRM tier join prevents it.

Do this NOW, not in week 1. Even if your "CRM" is a Google Sheet for the first month.

1. Open your CRM (or create a Google Sheet titled "tiered-contacts").
2. Add a column called "Tier."
3. Populate every contact you have at one of: - Tier 0 (4 to 8 people): family, co-founder, top 1 to 2 investors. Must never go unanswered for more than 24 hours. - Tier 1 (15 to 30 people): top investors, top customers, top advisors, board, key partners. Never more than 48 hours. - Tier 2 (50 to 200 people): active deals, warm prospects in motion, current customers - Tier 3 (everyone else): long-tail contacts, cold prospects, dormant
4. List the names of Tier 0 and Tier 1 explicitly in CLAUDE.md.

This step alone will catch 80% of the embarrassing missed-message failures.

Hour 2. Folder and three files

Open Cowork (Mac) or Claude Desktop (Windows). On the left, click "Add a folder." Name it [your-company]-os. Cowork creates a real folder on your computer.

Have Claude make the .md files for you

In Cowork, type this exact prompt:

"Create three files in this folder: CLAUDE.md, assistant.md, and tone-voice.md. Use the templates from Appendix F of the 'Build Your Own Sasha' playbook (paste below if I haven't already). After creating each file, ask me one question at a time to fill in the bracketed fields. Start with CLAUDE.md."

Paste in the templates from Appendix F if Claude doesn't have access to this doc.

Claude will create the file, then ask you one question at a time. Answer them. Takes about 25 minutes.

After the three files exist:

"Read CLAUDE.md, assistant.md, and tone-voice.md. Tell me in 5 bullets what you understand about me and how I work."

If the bullets sound generic, your inputs were thin. Iterate.

Hour 3. First connectors

Three things this hour: Calendar, Email, Telegram bot.

Google Cal: Cowork → Settings → Connectors → Browse → Google Calendar → Add → OAuth. Test:

"What's on my calendar tomorrow?"

Gmail: Settings → Connectors → Browse → Gmail → Add → OAuth. Test:

"Summarize my last 5 unread."

Outlook: Same path, pick "Microsoft Mail" / "Microsoft Calendar."

Telegram bot: 1. Open Telegram, search @BotFather, verify the blue checkmark. 2. Send /newbot. Name it Sasha for [Your Name]. Username ends in bot. 3. Copy the token. Treat like a password. 4. Message your bot any text to activate the chat. 5. In your browser: https://api.telegram.org/bot<YOUR_TOKEN>/getUpdates. Find "chat":{"id":...}. Copy the chat ID. 6. In Cowork:

"Save these to secrets.md in this folder. Telegram bot token: [PASTE]. Telegram chat ID: [PASTE]. Add a rule to CLAUDE.md: never read or output the contents of secrets.md unless I ask explicitly." 7. Test: "Send me a Telegram message that says 'Sasha is online' using secrets.md."

If your phone buzzes, the bot works.

Hour 4. Voice file (two-pass)

The voice file is the thing founders most often get wrong. Do not let the assistant average across your sent folder. Average gets you average.

Pass 1. You hand-pick five standout messages

Open your sent folder. Find five messages you're genuinely proud of. A range: - A short reply that closed a loop (under 30 words) - A medium email that landed (80 to 150 words) - A direct challenge to a peer that produced a good reply - A cold pitch that got a meeting - A passive recap email that captured the call

Paste each into tone-voice.md under "Verbatim samples I'm proud of." Annotate one line per sample explaining why it worked.

Pass 2. Assistant patterns around them

Now prompt:

"Read my five hand-picked samples in tone-voice.md. They represent my voice at its best. Now scan my last 50 sent emails (skip auto-routine, calendar accepts, forwards, and one-word replies). For each pattern in my samples (sentence shape, opener, closer, word choice), find supporting examples in the wider corpus. Update tone-voice.md to (a) confirm patterns with multiple examples, (b) flag patterns from the wider corpus that I should consider red lines, (c) propose 5 red-line phrases I never use. Show me the result. Do not save until I approve."

Iterate. Approve.

First draft test

Forward a real unanswered email to your assistant:

"Draft a reply in my voice using tone-voice.md. Under 100 words."

Mark every line that sounds wrong. Tell your assistant why. Re-draft. Capture each correction in /memory/lessons.md:

"Add to lessons.md: I never start with 'Hope you're well.' Default opener: 'Quick one,' or the verb of the ask."

By end of Hour 4 you have: - Working folder - Three populated .md files (CLAUDE.md, assistant.md, tone-voice.md) - CRM tier set up - Calendar + Email + Telegram bot live - Voice file built from your real best work - 5 to 10 lessons captured

Skip everything else. Don't try to do more in Pair Setup.

Stuff that breaks in Pair Setup. Top 6 errors.

Self-work. Four weeks. Weekly milestones.

Fifteen minutes a day. Specific weekly outputs.

What you do daily (every week)

1. Morning (2 min): Ask your assistant for your Power 3 and today's calendar.
2. After every meeting (3 min): Have your assistant pull the transcript, draft a recap in your voice, update tasks.md. Review and send.
3. After every voice correction (1 min): Add to lessons.md.
4. Before every cold message (2 min): Re-draft against tone-voice.md and lessons.md.
5. End of day (2 min): Wins, rollovers, tomorrow's Power 3.

Week-by-week milestones

If a week's milestone slips, do not advance. Fix that week before adding the next layer.

What you measure

If voice corrections don't drop in week 4, your tone-voice.md needs more samples. Re-run Pass 1.

Pair Tune-up session (4 hours, week 4)

Hour 1. Voice tuning

Read every entry in lessons.md from the last four weeks. Prompt:

"Read every lesson from the last four weeks. For each, propose an update to tone-voice.md or assistant.md so the rule is permanent. Show the diff."

Approve.

Hour 2. Walk through the three scenarios

(Setup details below in "The three scenarios.")

Hour 3. Schedule the cron tasks

Three scheduled tasks via Cowork's scheduled-tasks system:

Morning briefing. 8am weekdays.
Trigger: cron 0 8 * * 1-5
Prompt: "Read pre-meeting briefs for today. Pick Power 3 from tasks.md, energy-matched. Send to Telegram via secrets.md."

Inbox triage. Every 2 hours, 8am to 8pm.
Trigger: cron 0 8,10,12,14,16,18 * * 1-5
Prompt: "Run /skills/triage-loop.md."

Evening wind-down. 7pm.
Trigger: cron 0 19 * * *
Prompt: "Read tasks.md. Today's wins. Tomorrow's first meeting. Send to Telegram."

Add a heartbeat. Every scheduled task writes heartbeat.json with a timestamp at the end. Morning briefing flags any over 26 hours stale.

Three-tier scheduling resilience

Cowork's scheduled-tasks system is in research preview. Plan for it changing.

Most founders never need to leave Tier A. Have the failover plan documented in case Anthropic ships a breaking change.

Hour 4. Triage loop with CRM tier join

Update /skills/triage-loop.md to cross-reference every sender against the Tier field you set up in Pair Setup. Tier 0 and 1 always escalate, even if quiet.

By end of Pair Tune-up you have a working L3 system.

How memory works across surfaces (don't get burned by this)

Your assistant is one mental model. In reality, three different surfaces with three different memory rules.

If Sasha "forgot what we discussed yesterday," check which surface you used. The answer is almost always "different surface, different memory."

What never goes in (privacy red-lines)

Your assistant has access to your email, CRM, calendar, messaging, and secrets file. That data is logged per Anthropic's privacy policy with their retention windows.

These never go in your assistant's context:

For sensitive topics, open a fresh Cowork window with no connectors enabled and discuss verbally before writing anything down.

Vacation mode

Real CoS systems survive vacations. Define this before you go.

One-time setup

Create /skills/vacation-mode.md:

---
name: vacation-mode
trigger: "Sasha, vacation mode on" / "Sasha, vacation mode off"
---

When ON:
1. Auto-respond to incoming email (not from Tier 0 or 1) with: "Out until [date]. Will respond when back. - [Sign-off]."
2. Hold all triage drafts. Do not auto-send anything.
3. For Tier 0 and 1 messages, escalate via SMS to: [phone of designated backup human, e.g., spouse, EA, co-founder].
4. Pause morning briefing and evening wind-down.
5. Continue post-meeting recaps if I take any meetings during vacation.

When OFF:
1. Disable auto-respond.
2. Resume scheduled tasks.
3. Generate "while you were out" digest: every email, message, and meeting from the vacation period, grouped by Tier.
4. Telegram me the digest link.

Before you go

• Pick the designated backup human. Tell them.
• Test vacation mode for one day before the actual trip.
• Update CLAUDE.md "I'm offline from [date] to [date]" so every session knows.

When you're back

• Run "vacation mode off."
• Read the "while you were out" digest before opening any other tool.

The three scenarios (output first)

Scenario 1. Pre-meeting brief and post-meeting recap

What you get

Every weekday at 8am. Telegram message:

"Three external meetings today. Briefs ready. Highest priority: [name] at 10am, who you last met at GEM Getaway in April. Two open promises from your side."

Click the link. A folder of [YYYY-MM-DD]-[name].md files. Each one has last touchpoint, open promises, three suggested questions, one warm-context line, sensitivity flags.

Within 30 minutes of every meeting ending. Telegram nudge with a draft recap in your drafts folder. Tasks.md updated. CRM updated. People profile updated.

Voice and WhatsApp calls work the same way. Plaud transcribes. Skill 2 runs.

What you check (your checklist)

Pre-meeting: - [ ] Brief arrived in Telegram by 8am - [ ] Brief includes open promises - [ ] Brief includes three useful questions - [ ] Sensitivity flags raised if relevant - [ ] Dial-in / location is right

During: - [ ] Recorder on (Granola for digital, Plaud for in-person/phone) - [ ] Talking, not typing - [ ] Confirmed next steps out loud before ending

Post-meeting (within 30 min): - [ ] Recap draft arrived in Telegram - [ ] Recap sounds like you wrote it - [ ] Action items right - [ ] Recap sent - [ ] CRM updated correctly

How it's set up

Skills /skills/pre-meeting-brief.md and /skills/post-meeting-recap.md (Appendix G). Pre-meeting wired to a 9pm-the-night-before scheduled task. Post-meeting triggers when Granola/Plaud lands a transcript or when you say "I just got off a call with [name]."

Scenario 2. Cross-channel triage

What you get

Every 2 hours during work hours. Telegram nudge:

"8 messages need eyes. 2 drafts ready. 1 needs your judgment: [name] re: [topic]."

Click through. A feed at /scan-reports/[YYYY-MM-DD-HHmm].md with sender, channel, time, summary, suggested action (REPLY/WAIT/IGNORE/ESCALATE), and drafts for REPLY items.

You approve drafts. Send. 7 minutes for what used to be 90.

What you check

• [ ] Triage nudge arrived on schedule
• [ ] Every Tier 0 and 1 contact represented
• [ ] LinkedIn DM caught (not just email)
• [ ] Suggested actions reasonable
• [ ] Drafts in your voice
• [ ] Approved and sent

How it's set up

/skills/triage-loop.md runs every 2 hours via a scheduled task. Pulls email, Beeper (iMessage/WhatsApp/LinkedIn/IG/SMS), Slack, CRM. Joins every sender against Tier. Tier overrides volume.

Scenario 3. Conference + Plaud sub-playbook

Full setup in Appendix K.

What you get

• Three weeks before: a folder of attendee intel briefs with verbatim quotes from prior touchpoints
• During: real-time Plaud-to-attendee tagging via your voice memos
• Day +1 to +3: cleaned per-attendee transcripts with diarization fixed and Plaud errors flagged
• Day +3 to +7: draft first-touch follow-up per starred attendee in your voice

You review every draft. AI does not auto-send first-touch.

What else is possible (level-up menu)

After the three scenarios are stable:

• Travel coordination. Flights and hotels based on calendar and your loyalty programs.
• Personal admin. Doctor appointments, family logistics, school pickups, house projects.
• Investment / portfolio updates. Weekly summary with macro context.
• Content drafting. LinkedIn, X threads, blogs in your voice.
• Press / podcast prep. Briefs based on host's recent work.
• Daily wins log. End-of-day capture surfaced in weekly review.
• Team 1:1 prep. Briefs for every direct report based on their last week.
• Inbound voice line. ElevenLabs + Vapi for an AI front desk.

Each is one new /skills/[name].md file. About an hour of setup each.

Six failure modes. Six recoveries.

Pricing and how to engage

Two paths. Both start with the same playbook.

Why $25,000 for the day, $295,000 for the org.

$25K is less than a month of analyst headcount. The outputs (morning brief, post-meeting recap drafts, triage feed) save a working founder 50+ hours a month. Pay-back is measured in weeks, not quarters. Anchored against "the analyst you almost hired," not against "the SaaS tool you almost subscribed to."

$295K is what it costs to drag a real consulting firm into a 3-month change program for one function. We do it for the whole exec floor, with the actual artifacts (memory files, voice files, scheduled tasks, IT posture) that the team owns at the end. No retainer trap.

What's included at $25,000:

• Eight hours, in person, on your turf, anywhere in the US
• Travel and hotel covered
• Your voice file locked by end of day
• Tier 0 and Tier 1 names approved by end of day
• Telegram bot wired, scheduled tasks live
• One follow-up tune-up call within 30 days

What's included at $295,000:

• Up to 10 exec leaders, each with a working CoS
• Per-leader voice training, individual to each exec
• Connector standardization across the team
• IT and compliance playbook written for your CISO
• Weekly office hours for 12 weeks
• Custom skill library per function (sales, finance, product, ops)

What's not included (either tier):

• Hardware (Plaud, AirPods, etc.). Passed through at cost.
• Software subscriptions (Claude Max per seat, Beeper, Granola, etc.). Passed through at cost.
• Ongoing managed service. You own the artifacts at the end. If you want a managed tier later, separate conversation.

How to engage: email ceo@100x.inc. For the day, name your city and three candidate dates. For the org rollout, name your team size and current stack. Nik replies within two business days.

Closing. The 24-hour next action.

Most founders read a playbook like this and don't start.

Don't be one. The next 24 hours, do four things:

1. Download Claude Cowork (or Claude Desktop on Windows). 10 minutes.
2. Block 4 hours on your calendar this Saturday morning. Title it "Pair Setup: build my Sasha." Invite a buddy if you have one.
3. Install Beeper. Log in to your messaging channels inside Beeper. 20 minutes.
4. Order Plaud Note only if your meeting mix is heavy in-person or phone. Otherwise just install Granola.

That's it. The decision tree starts there.

In five weeks you have the version that drafts your morning, recaps your meetings, and triages your inbox.

Either you build it or someone faster does.

- Nikolai

---

APPENDICES

Appendix A. The architecture. Four boxes. One brain.

┌───────────────────────┐     ┌──────────────────────┐
│  CAPTURE              │     │  DESKTOP BRAIN       │
│                       │     │                      │
│  Voice notes (Plaud)  │ ──▶ │  Claude Cowork (Mac) │
│  Digital meetings     │     │  or Claude Desktop   │
│  (Granola)            │     │  (Windows)           │
│  Email                │     │                      │
│  Messaging (Beeper)   │     │  Reads CLAUDE.md     │
│  Internal team chat   │     │  Reads tasks file    │
│  (Slack / Roam)       │     │  Calls connectors    │
└───────────────────────┘     └──────────────────────┘
                                        │
                                        ▼
                              ┌──────────────────────┐
                              │  MEMORY              │
                              │                      │
                              │  CLAUDE.md           │
                              │  /memory/            │
                              │  /accounts/          │
                              │  /tasks.md           │
                              └──────────────────────┘
                                        │
                                        ▼
                              ┌──────────────────────┐
                              │  ACT                 │
                              │                      │
                              │  Email drafts        │
                              │  Calendar holds      │
                              │  CRM updates         │
                              │  Telegram nudges     │
                              │  Doc creation        │
                              └──────────────────────┘

No backend. No database. No custom app. Claude, a folder, the right connectors.

Appendix B. The full MECE connector menu

B1. Calendar

B2. Email

B3. Direct messaging (Beeper Desktop today)

Beeper bridges: iMessage, WhatsApp, LinkedIn DMs, Instagram, Matrix, SMS, Telegram, Discord, Signal.

X DMs: currently unsolved (post-2024 API lockdown).

Failover plan documented in Step 0.

To install Beeper MCP in Cowork: Settings → Connectors → Beeper Desktop → Add. Beeper Desktop must be running.

B4. Internal team chat

B5. CRM

B6. Documents and notes

B7. Voice meeting transcripts

B8. Notification channel

B9. Desktop / mobile control (advanced)

B10. Voice agent (L4+, optional)

B11. Human routing (optional)

A real EA or CoS, with rules for what is escalated. See "AI vs human EA vs both" at the top of this doc.

Appendix C. Folder structure

[your-company]-os/
├── CLAUDE.md
├── README.md
├── assistant.md
├── tone-voice.md
├── tasks.md
├── tasks-private.md
├── secrets.md                # Never read aloud
├── /memory/
│   ├── people/
│   ├── projects/
│   ├── lessons.md
│   └── archive/
├── /accounts/
│   └── _template/
├── /channels/
│   ├── email-templates/
│   ├── slack-templates/
│   └── disclaimers.md
├── /skills/
│   ├── pre-meeting-brief.md
│   ├── post-meeting-recap.md
│   ├── triage-loop.md
│   └── vacation-mode.md
├── /scheduled-tasks/
├── /scan-reports/
├── /conferences/
└── /design/

Appendix D. The agent maturity matrix (LB to L5)

Borrowed from self-driving cars. Who's driving, when does the human let go?

LB. The Black Hole. No process. Work disappears. L0. Parked Car. Process exists. 100% manual. L1. Cruise Control. AI handles one narrow task. L2. Adaptive Cruise + Auto-Brake. Chain of tasks in one channel. L3. Lane Change Assist. Multi-step across 2 to 3 systems. L4. Full Self-Driving (Supervised). End-to-end. Human reviews weekly. L5. Waymo. Fully autonomous. Self-improving.

Six dimensions: autonomy, context, decision-making, error handling, learning, human role. Overall level = lowest score.

Most "AI agents" in production are L1.

After Pair Setup + 4 weeks + Pair Tune-up, you're at L3. Climbing to L4 takes another 3 to 6 months.

Full matrix: 100 Claude OS/projects/active/agent-maturity-matrix/agent-maturity-matrix.md.

Appendix E. The skateboard model (incremental delivery)

MVP gets misread as "build a wheel, then a chassis, then a car." Three of those four are useless on their own.

The right read, per Henrik Kniberg's Making Sense of MVP: at every step, ship the smallest usable version of the whole thing. A skateboard. A scooter. A bike. A motorbike. A car.

This playbook follows that. Pair Setup ships a skateboard (working brief + voice + Telegram nudge). Self-work weeks 1 to 4 ride the skateboard, then upgrade. Pair Tune-up ships the bike (full triage + scheduled tasks).

Appendix F. The three templates

F1. CLAUDE.md

# Memory
**Owner:** [Your Name] | **Version:** 1.0 | [Today's date]

## Me
[Your name], [your role] at [your company].
[What you build. What you do.]
LinkedIn: [link]
Email: [primary work email]
Phone: [number]
Philosophy: [one line]
Constraints: [ADHD? Time zones?]

## My Assistant
**"Hey [AssistantName]"** = invoke the full system. Claude IS [AssistantName] when working in this workspace.

| Component | Where |
|-----------|-------|
| Tasks | `tasks.md` |
| Private tasks | `tasks-private.md` |
| Voice training | `tone-voice.md` |
| Persona | `assistant.md` |
| Daily scheduled tasks | `/scheduled-tasks/` |
| People notes | `/memory/people/` |
| Lessons | `/memory/lessons.md` |
| Secrets | `secrets.md` (never read aloud) |

## Operating Rules
1. Files govern behavior.
2. First action every session: read CLAUDE.md, lessons.md, tasks.md.
3. Routing table below for where files go.
4. Label Facts / Assumptions / Judgment in analysis.
5. No em-dashes in outbound drafts.
6. Sign off as [Sign-off Name] with AI disclaimer per tier.
7. Never read or output secrets.md unless I ask explicitly.
8. Confirm before generating. Never invent named-mutual claims or customer status.
9. Never put PHI, banking credentials, M&A non-public info, attorney-privileged content, HR investigation details, or insider info in any file.

## Routing Table
| Type | Where |
|------|-------|
| Notes about a person | `/memory/people/[firstname-lastname].md` |
| Project plan | `/memory/projects/[project-slug].md` |
| Meeting recap | `/accounts/[slug]/recaps/[YYYY-MM-DD].md` |
| Outbound draft | `/channels/email-templates/` |
| Lesson | `/memory/lessons.md` |

## People (with Tier)
| Who | Role | Channel | Tier |
|-----|------|---------|------|
| [name] | [role] | [channel] | [0-3] |

## Projects (active)
- [Project A]: [one line]
- [Project B]: [one line]

## Working Hours
- Working hours: [9 to 5 PT]
- Deep work: [9 to 1]
- Admin block: [1 to 3]
- Channels I check: [email, Slack, iMessage]
- Channels I never check: [LinkedIn DMs, Discord]

## Mandatory Rules
**Archive Rule:** Before overwriting any CLAUDE.md or memory file, archive to `/memory/archive/{filename}_{YYYY-MM-DD}_{description}.md`.
**Existing Work Check:** Before creating, search `/memory/` and `/accounts/` for prior art.
**Reference Integrity:** When renaming, grep for old path everywhere and update.

## Who [AssistantName] Is
[AssistantName] is not a chatbot. Not a copilot.
[AssistantName] is the [role] for [Your Name] at [Your Company].
[AssistantName] runs the scaffolding so [Your Name] keeps the judgment work.

## Vision
"[One sentence on what [AssistantName] is here to do.]"

## How [AssistantName] Speaks
- Short sentences. Every word earns its place.
- Action first.
- No em-dashes.
- Treats the recipient as an equal adult.
- One question per message.

## What Triggers a Hand-back to [Your Name]
- Anything involving money over $[threshold].
- Anything involving a person flagged sensitive in /memory/people/.
- Anything legal, HR, or a customer escalation.
- Anything covered by the privacy red-lines in CLAUDE.md.
- Anything where [AssistantName] is less than 80% confident.

## AI Disclaimer (mandatory)
Tier 1 (under 50 words, casual): no disclaimer if recipient already knows.
Tier 2 (50 to 200 words, semi-formal): "Drafted by my AI chief of staff. Reviewed by me."
Tier 3 (200+ words, formal): "This was drafted by my AI assistant. Anything I missed, please push back."

## Sign-off
Every outbound message ends with: `- [Sign-off Name]`

F3. tone-voice.md (most filled in by your assistant in two-pass)

# [Your Name]: Tone and Voice
**Version:** 1.0 | [Today's date]

## How I Write

### Length
- iMessage / WhatsApp: [8 to 35 words]
- Slack: [1 to 3 sentences]
- Email: [80 to 150 words, three paragraphs max]
- LinkedIn DM: [30 to 80 words]

### Sentence shape
- [Subject. Verb. Object.]
- [Short beats compound.]
- [Full stops. Never em-dashes.]

### Word choice
- [Plain English.]
- [No jargon unless they used it first.]
- ["Yes" or "no" before reasoning.]

### Opener defaults
- [Auto-filled by your assistant, Pass 2]

### Closer defaults
- [Auto-filled by your assistant, Pass 2]

## Verbatim samples I'm proud of (Pass 1: hand-picked by me)

### A short reply that closed a loop
> [Paste a real one. Annotate why it worked.]

### A medium email that landed
> [Paste a real one. Annotate.]

### A direct challenge to a peer that produced a good reply
> [Paste a real one. Annotate.]

### A cold pitch that got a meeting
> [Paste a real one. Annotate.]

### A passive recap email that captured the call
> [Paste a real one. Annotate.]

## Supporting patterns (Pass 2: assistant fills in)
- [Auto-filled]

## Red lines (Pass 2: assistant fills in, I review)
- [Auto-filled]

Appendix G. The four core skill files

G1. /skills/pre-meeting-brief.md

---
name: pre-meeting-brief
trigger: Scheduled task at 9pm OR "Hey [AssistantName], brief me on my [time] meeting"
---

For every external meeting tomorrow:
1. Pull contact profile from CRM.
2. Pull last 30 days of email both directions.
3. Pull last 30 days of iMessage, WhatsApp, LinkedIn DMs via Beeper.
4. Pull the last meeting transcript (Granola or Plaud).
5. Pull open promises (mine and theirs) from tasks.md.
6. Write brief.md to /accounts/[slug]/briefs/[YYYY-MM-DD].md with:
   - Last touchpoint and what was said
   - Open promises
   - Three suggested questions
   - One line of warm context
   - Sensitivity flags
7. Send Telegram ping at 8am: "Briefs ready. Highest-priority: [name] at [time]."

G2. /skills/post-meeting-recap.md

---
name: post-meeting-recap
trigger: Within 30 minutes of meeting end
---

1. Pull transcript from Granola or Plaud.
2. Extract action items.
   - Mine, add to tasks.md.
   - Theirs, add to "WAITING ON".
3. Extract decisions to /accounts/[slug]/decisions.md.
4. Update /memory/people/[name].md.
5. Draft recap email. 90 to 150 words. My voice. No em-dashes. Stage in drafts.
6. Update CRM:
   - Log call activity.
   - Update deal stage if signals warrant.
   - Add new tasks to deal record.
7. Telegram ping: "Recap drafted for [name]. Review and send. CRM updated."

G3. /skills/triage-loop.md

---
name: triage-loop
trigger: Scheduled task every 2 hours during work hours
---

1. Pull unread + recent activity from email, Beeper (iMessage/WhatsApp/LinkedIn/IG/SMS), Slack, and CRM tasks.
2. Cross-reference every sender against CRM Tier field. Tier overrides volume.
3. Filter to Tier 0, 1, 2 only. Drop routine notifications.
4. For each remaining item, write triage row:
   - Sender, channel, time, one-line summary
   - Suggested action: REPLY / WAIT / IGNORE / ESCALATE
   - For REPLY items, draft a response in my voice
5. Save feed to /scan-reports/[YYYY-MM-DD-HHmm].md.
6. Telegram summary: "[N] messages need eyes. [N] drafts ready. [N] need your judgment."

G4. /skills/vacation-mode.md

(Full template in "Vacation mode" section above.)

Skill file template (for any new skill)

---
name: [skill-name]
description: [one sentence]
inputs: [list]
outputs: [list]
---

## Trigger
[How it fires.]

## Steps
1. [step]

## Failure modes
- [if X, do Y]

## Voice
[Output for me, or output for the recipient?]

## Version
1.0 / initial

Appendix H. Mac vs Windows

Same on both. Cowork (Mac) and Claude Desktop (Windows) read CLAUDE.md identically. Beeper Desktop runs on both. Cloud connectors are identical. Scheduled tasks are identical.

Mac wins at. Native iMessage via SQLite chat.db. osascript bridge to Apple Notes, Mail, Calendar, Reminders, Finder. Native mic capture. "Control your Mac" MCP for full desktop automation.

Windows wins at. PowerShell scripting. AutoHotkey for any UI macro. Native Outlook integration is cleaner.

Cross-platform fallback. Windows + iMessage? Use Beeper. Mac + Outlook? Use Microsoft Graph.

Appendix I. Avatar

Your assistant should have a face. Trust is built faster on something visualized.

Reference assets in this playbook: - assets/sasha-avatar.png: Sasha headshot - assets/viv-sasha-pair.png: Sasha next to her sibling agent Viv - assets/wardrobe-midshots-v2/: 12 mid-shots for dashboards, emails, status pages

Build yours: 1. Write a one-paragraph character seed: name, age, role, vibe, signature outfit. 2. Generate variations in Gemini Imagen, Midjourney, or Grok. 12 mid-shots, same face, varied wardrobe. 3. Pick one canonical. Reference her in every dashboard. Consistency builds trust.

Appendix J. The reference implementation (the live Sasha)

Sasha is real. She runs Nik Shah's CEO workflow at 100x. Live since March 2026. Currently L3 climbing to L4.

Source paths: - Architecture: 100 Claude OS/projects/active/sasha/ARCHITECTURE.md - Tasks: 100 Claude OS/projects/active/sasha/TASKS.md - Scheduled tasks: sasha/design/scheduled-tasks-prompts.md - Outbound voice (Nikolai): 100 Claude OS/projects/active/nikolai/ - Lessons learned: 00 Claude CEO/memory/lessons.md

Sasha's hardest lessons (skip them): 1. Voice samples beat voice rules. Hand-pick your best, then have the assistant pattern. 2. Heartbeats prevent silent failures. 3. Tier joins prevent missed Tier-1 messages. 4. AI disclaimer rules must exist before auto-send. 5. Plaud transcripts are 10x more useful with attendee-list cross-reference. 6. iMessage chat.db has an attributedBody column most pipelines miss. 7. Confirm before generating named-mutual claims.

Appendix K. Conference + Plaud sub-playbook

Stand-alone playbook for the most under-leveraged founder workflow.

K1. Hardware

K2. Three weeks before

Drop attendee CSV in /conferences/[slug]-YYYY-MM/attendees.csv. Columns: name, company, title, email if known.

"Pull intel on every attendee. For each: LinkedIn role, company news, my last touchpoint across email, iMessage, WhatsApp, LinkedIn, CRM. Verbatim quotes preferred over summaries. Flag people I've never met. One MD per person in /conferences/[slug]-YYYY-MM/intel/."

Star 5 highest-priority. For each starred:

"Write a one-line conversation opener mapped to a real value-offer for [name], based on what they're working on right now."

K3. The day before

"Build a per-day attendee plan. For each conference day, list which sessions starred attendees are attending, suggested 5-minute conversations, and a one-line opener. Save to /conferences/[slug]-YYYY-MM/day-plan.md. Telegram me the link tomorrow morning."

K4. During the conference

Plaud clipped to lanyard. Always on.

After each meaningful conversation, dictate 30-second voice memo:

"Just talked to [name] from [company]. We discussed [topic]. They want [thing]. Next step is [thing]."

Drop voice memos into /conferences/[slug]-YYYY-MM/voice-tags/.

End each day:

"Pull today's Plaud transcripts. Cross-reference my voice tags. For each starred attendee I touched today, write a one-paragraph summary. Save to /conferences/[slug]-YYYY-MM/daily-summaries/[YYYY-MM-DD].md."

K5. Day +1 to +3 (cleanup)

"Cross-reference attendee list against all Plaud transcripts. Replace 'Speaker 1' with real names. Flag every proper noun that looks like a Plaud transcription error (verify against attendee list). Extract verbatim quotes per attendee with timestamps. One cleaned MD per attendee in /conferences/[slug]-YYYY-MM/cleaned/."

Review the cleaned folder. Mark errors the assistant missed.

K6. Day +3 to +7 (follow-ups)

"Draft a first-touch follow-up message in my voice for [name]. Reference what we discussed (use verbatim quotes from cleaned/[name].md). 8 to 50 words. Specific value-offer mapped to what they want. Use iMessage register (peer, casual) unless their age or vibe suggests email. Sign off as [Sign-off Name]. Stage drafts in /conferences/[slug]-YYYY-MM/follow-ups/[name]/."

Review every draft. AI does not auto-send first-touch.

K7. Conference user checklist

Three weeks before: - [ ] Attendee CSV uploaded - [ ] Intel folder generated - [ ] 5 starred targets picked - [ ] One-line openers written for each - [ ] Plaud charged, lanyard ready

Day before: - [ ] Day plan generated - [ ] Tomorrow's Telegram nudge confirmed - [ ] Power banks packed

During: - [ ] Plaud running - [ ] Voice tag after every meaningful conversation - [ ] Daily summary generated end of day

Day +1 to +3: - [ ] Plaud transcripts pulled - [ ] Cleanup ran across attendee list - [ ] Per-attendee MDs reviewed for errors

Day +3 to +7: - [ ] First-touch drafts staged for every starred attendee - [ ] Drafts reviewed and sent (one at a time) - [ ] CRM updated with conference touch on every starred contact

K8. Conference-specific failure modes

Appendix L. Agent naming guide (compressed)

Full version: 100 Claude OS/projects/active/agent-maturity-matrix/agent-naming-guide.md.

The 8 laws:

1. Identity over history. Name encodes what the agent IS, not where it came from.
2. One or two syllables. Three only if already proven.
3. Alliteration with function. Sasha ↔ Shreya, Sophei ↔ HEI, Faz ↔ Finance.
4. Posh and elite. Sounds right at a private members' club.
5. Effortless pronunciation. Hey-Siri test. No silent letters for product brands.
6. Absolute uniqueness. Google "[name] AI". If a known product owns it, reject.
7. American/British roots, or South Asian bilingual lane. Lives at home in NYC, London, SF.
8. One name, one agent. No sharing.

Source

• 100x agent maturity matrix
• 100x agent naming guide
• Henrik Kniberg, Making Sense of MVP
• The live Sasha
• The live Nikolai
• 32+ lessons in 00 Claude CEO/memory/lessons.md

Send corrections to nik@100x.inc. They land in /memory/lessons.md and propagate to the next version.

- Nikolai