Conference Talk

How a CMO
Vibe-Coded an
AI Chief of Staff

Denis Elkin

CMO, Alps2Alps

MIPT · 15 years in marketing · Hegai since 2020 · last coded 15 years ago

This presentation is being vibe-coded right now

What I'm NOT talking about

AI in the company

Cursor, Claude for team, internal tools

AI in marketing

Landings, video, creatives, analytics, research

AI in programming

Copilot, code generation, debugging

Today = only my personal AI assistant

One person's system. Built by that person. No engineering team.

The problem

Monday, peak season. 47 emails, 12 Slack threads, 26 Asana tasks, 2 project launches this week.

Two hours later I haven't started what I sat down to do.

The noise

● 2 email inboxes + Slack + Asana
● Telegram + channels + meetings
● 10 people, 5+ projects, 350K passengers

Frameworks I know but can't apply

● Dorofeev (task management)
● 4DX (execution discipline)
● GTD, Deep Work, etc.

You know it. You can't do it in the noise.

v0

ChatGPT

"What are you working on?"

We all went through this. Copy-paste context every session. Lost prompts. Zero state. Starts from scratch every time.

v1

Cursor + shared memory

Heard Seva Ustinov. Set up Cursor with CLAUDE.md — a shared memory file the AI reads every session.

Pros

+ Builds fast
+ MCP tools
+ Persistent context

Cons

- Laptop closed = gone
- No phone access
- No automation

v2

Telegram bot

● Voice message → transcribed (Grok, free) → categorized → actionable task
● Email draft in 10 seconds
● 5,200+ lines of code
● First "holy shit" moment

Accessible from phone. Works 24/7. Background schedulers.

v3

Full system

All three interfaces connected to the same brain.

77

tools

19

schedulers

20

skills

2

servers

8-page React dashboard · unified inbox · team · finance · meetings

THE differentiator

Thinking partner

11 frameworks from real books. AI asks questions, doesn't give answers.

"INTERVIEW BEFORE OUTPUT — never give recommendations until you've interviewed Denis."

Dorofeev 4DX Descartes Square Stakeholder Sim Board Simulation Premortem

How it thinks

Not a chatbot. A chief of staff.

Deep context

Knows my email rules, my team, my frameworks, my voice. Reads goals and open questions every session.

Plan-critique

3 roles (architect, PM, domain stakeholder) attack every plan silently before it reaches me.

Stress-test

Every output attacked in 3 rounds before presenting. Testable claims = tested, not reasoned.

Security & control

Auth proxy on VPC. Tests + multi-role code review before every deploy. Never sends without approval.

Failures & cost

Voice-to-text → hallucination chain

Wispr Flow misrecognized names. Wrong names hit the prompt → AI hallucinated on top. Caught it before it went out.

AI deleted its own API keys

Cursor overwrote .env while editing config. System went dark. Lesson: .env is now read-only.

$$$: Runaway API calls

System decided to parse my personal inbox via API. Found out when credit card charge notifications started flooding my phone.

Monthly cost

Claude Max + $10-20/mo servers

Cheat codes

claude -p on a server

Claude Max ($100-200/mo) includes unlimited CLI use. Run claude -p "prompt" on your server instead of paying per API token. Saves thousands.

Wispr Flow — talk, don't type

We speak 3-5x faster than we type. Wispr Flow transcribes speech directly into any text field. Press Fn, dictate, done. Prompts, emails, tasks — all by voice.

I build AI systems across different functions in the company.
If you're doing the same or have similar challenges —
let's sync up.

Telegram: @denyolkin

Channel: @denis_AMM

LinkedIn: /in/denyolkin

Email: denis.elkin@amitours.com

How a CMOVibe-Coded anAI Chief of Staff