From Cone Red's AI-First Transformation Practice · v1 · 2026

Building an
AI-first company.

How to turn a team into a compounding
knowledge organism.

Dima Levin

Practice Lead · Cone Red AI-First Transformation Practice

38 slides · ≈ 25 min · skim in 5

2026-04-27

Not by adding AI
—
by re-architecting around it.

If you only watch 5 slides

A 25-minute deck. Five slides if you're skimming.

SLIDE 05

The 3 things AI is good at

Why "AI-first" works at all.

SLIDE 12

The 4-layer architecture

Vault · schema · agents · graph.

SLIDE 21

90-min meeting → 20-min tickets

The math everyone remembers.

SLIDE 35

"It's an operating system"

The thesis in five words.

SLIDE 36

Three things to do this week

If you want it live by Friday.

Read this with a CFO who has 5 minutes. Read the full deck if you're going to ship it.

01 — Section

The thesis.

Most companies have added AI.

Few are AI-first.

42% of organizations scrapped most of their AI initiatives in 2025 — up from 17% in 2024.

The fastest-deteriorating spend category in the enterprise.
Source: S&P Global Voice of the Enterprise, n=1,006 (Oct 2025).

What we hear from inside Fortune 500 right now

Anonymized — a Fortune 500 international food-and-beverage conglomerate. Employees reached out directly, bypassing their corporate AI rollout.

"Our monthly KPI is the number of prompts we run. Not outcomes."
"We're mandated to use Copilot — even though we're non-tech and it's the wrong tool."
"We don't know what other AI tools exist."
"We don't know how to use them."
"We don't know which problems they should solve."

Three weeks ago, The Information reported Meta did the same publicly: an 85,000-person "Claudeonomics" leaderboard. Pulled in two days. McKinsey/BCG/Deloitte cannot name this dysfunction. We can.

Why AI-first works

AI is genuinely good at three things.

Text — markdown, frontmatter, plain English
Conversation — colleague, not search engine
Memory — compounds when memory is durable

Your operating model is probably hostile to all three.

02 — Section

The four shifts.

What changes, in plain English.

Shift 1

From SaaS-as-truth...

...to vault-as-truth.

One git repository owns everything that matters. Markdown for the body, YAML frontmatter for the structure, git for the history. SaaS apps become render layers.

Shift 2

From hand-authored work...

...to AI-drafted, human-reviewed.

The bottleneck isn't thinking — it's transcription. Three agents (transcript extractor → product-manager → ticket generator) read what humans read and draft what humans then review. Humans edit, never start blank.

Shift 3

From documents-as-files...

...to documents-as-graph.

Every document is a node. /graphify walks the repo and produces a navigable map of how your company actually thinks. Knowledge becomes inspectable.

Shift 4

From individual heroics...

...to compounding institutional memory.

Knowledge lives in the vault, indexed in the graph, queryable by any teammate. When the hero leaves, intelligence stays.

03 — Section

Architecture.

Four layers. Each one cheap to install, expensive to skip.

The four-layer model.

flowchart TB subgraph L1 [" "] V[("📁 Vault — md + git")] end subgraph L2 [" "] S["📐 Schema — frontmatter conventions"] end subgraph L3 [" "] A["🧭 Agents — strategy · product · design · transcripts · negotiation"] end subgraph L4 [" "] G["🕸 Graph + 🧠 Memory"] end V --> S --> A A -.-> V V --> G style V fill:#FF4D4D,stroke:#FF4D4D,color:#fff,stroke-width:2px style S fill:#1A1A1A,stroke:#FF4D4D,color:#E8E8E5,stroke-width:2px style A fill:#1A1A1A,stroke:#FF4D4D,color:#E8E8E5,stroke-width:2px style G fill:#1A1A1A,stroke:#D4A574,color:#E8E8E5,stroke-width:2px

Layer 1

The Vault.

One git repository. Every document, ticket, transcript, decision. Plain markdown.

Version history. Branching. No vendor. The repo is yours.

Layer 2

The Schema.

Markdown without conventions is a swamp. Markdown with consistent YAML frontmatter is a database.

Pre-commit hook validates every change. Invalid output blocks the commit.

Layer 3

The Agents.

Each role on the team gets an agent counterpart — a persona with codified principles, frameworks, and outputs. Not chatbots; reusable thinking partners.

Strategy agent

McKinsey-grade frameworks · red-team · review

Product-manager agent

Tickets · estimates · roadmap synthesis

Design-critique agent

Visual direction · taste · adversarial eye

Transcript-extraction agent

Pains · decisions · BANT · action items

Negotiation-prep agent

Three-lens stress test · concession map

Presales agent

Proposals · win-strategy maps · clarification

Layer 4

The Graph + Memory.

After six months, you can't read all of the vault. Two systems make sure that's fine.

🕸 Knowledge Graph

Built by /graphify. Nodes, edges, communities. Reveals clusters you didn't know you had.

🧠 Cross-project Memory

At ~/.knowledge/. Persistent across sessions. Bootstrap once with /init-memory.

04 — Section

The init-* family.

Five commands. ~90 minutes. Your OS is up.

Five commands. One operating system.

/init-os

CLAUDE.md · BOARD · JOURNAL · DECISIONS · commands · agents · settings.

/init-knowledge

Wires this project into the cross-project knowledge hub at ~/.knowledge/.

/init-memory

Durable memory layer. Survives across sessions. Knows you tomorrow.

/init-agents

Audit + install role agents from the central library — strategy, product, design, transcripts. Keeps them current.

/graphify

Walks the vault, produces a navigable knowledge graph. First run is a revelation.

Run in this order. Idempotent — re-runs upgrade gaps.

05 — Section

The AI-PM pipeline.

Tickets stop being typed. They start being drafted.

Three agents. One human. One safety net.

flowchart LR T["📜 Transcript"] --> C["🎤 Transcript agent
extract"] C --> P["📝 PM agent
synthesize"] P --> V["📁 Ticket generator
write to vault"] V --> H["🛡 Pre-commit hook
validate"] H --> R["👤 Human
review · edit · commit"] style T fill:#1A1A1A,stroke:#71757A,color:#E8E8E5 style C fill:#1A1A1A,stroke:#FF4D4D,color:#E8E8E5 style P fill:#1A1A1A,stroke:#FF4D4D,color:#E8E8E5 style V fill:#1A1A1A,stroke:#FF4D4D,color:#E8E8E5 style H fill:#1A1A1A,stroke:#D4A574,color:#E8E8E5 style R fill:#FF4D4D,stroke:#FF4D4D,color:#fff

The math

A 90-min meeting
= 5 to 10 tickets.

By hand

~3 hrs

Tired PM at 5pm.
Details lost. Quality variable.

By pipeline

~20 min

5 min agents · 15 min review.
Every line cited.

Hand-authoring: default → exception.

06 — Section

This week.

The only week with friction. Get through it.

Day by day.

Mon

Install Claude Code

Everyone. Tech AND non-tech. 30 min max.

Mon

Clone the repo

Three commands: clone · cd · claude.

Tue

Read CLAUDE.md

30 min. Then quiz Claude on what you read.

Wed

Co-work · 90 min

I run the init-* family live. You watch + ask.

Thu

First real task

Pick something real. Do it in Claude Code.

Fri

/journal + retro

30 min. What worked. What was clunky.

For engineers

The tech path.

# install (one time)
npm install -g @anthropic-ai/claude-code

# auth (one time)
claude login

# every day
cd company-os
claude
> /whoami
> /journal
> /task init T-042

Branches per work. PRs for shared files. Merge to main when reviewed.

For everyone else

The non-tech path.

You're not coding. You're chatting.

cd ~/Desktop/company-os
claude

> What's the latest on Tampere?
> Summarize Tuesday's call.
> What did we decide about pricing last quarter?
> Draft an email to Mike.

Stop searching. Start asking.

Wednesday

90 minutes.
Screens shared.

No homework. No prep. Bring one real question you'd normally ask in Slack.

Most non-tech teammates flip the switch in this session.

07 — Section

Culture.

Trees · PRs · Merges · Journal.

Four primitives. Old behaviors. New venue.

🌳

Trees / branches

Your private thinking space. Cheap to create, cheaper to throw away.

🔀

PRs / conversations

Not paperwork. Reviewers catch what you missed. 24-hour turnaround.

✅

Merges / commits to the future

Your future self reads commit messages. Be intentional.

📓

Journal / company diary

Telegram is now-talk. Journal is forever-talk. Don't confuse them.

"Tools shape behavior. Behaviors shape culture.
The behaviors that make AI-first work are old:
write things down, review each other's work, leave a trail."

08 — Section

The first 90 days.

Milestones to expect.

From friction to compounding.

1

Week 1 — Friction week

Install · clone · co-work. Friction highest now. Friday retro.

2

Week 2 — Habit-forming

Daily journal becomes routine. First non-tech PRs. No one opens ClickUp out of habit.

3

Month 1 — First /graphify

The team "sees itself." Use the graph to decide what's next.

4

Month 2 — AI-PM lights up

80% of new tickets from the pipeline. 20% hand-authored (the truly novel work).

5

Month 3 — First compounding win

A new project lands. Cross-project synthesis from ~/.knowledge/ skips a week of discovery. Mark the date.

09 — Section

Failure modes.

What kills this if you're not paying attention.

Six anti-patterns.

Treating Claude Code as a search engine — talk to it, don't query it
Letting AI write but not reason — use it for thinking, not just typing
Skipping the schema — pre-commit hook on Day 1, no exceptions
One person owns the OS — bus factor > 1, always
Over-engineering before the team uses it — ship MVP, add layers later
Letting Slack/Telegram eat the journal — move it or lose it

Avoid these six and you'll succeed. Don't, and the failure will look like "AI just doesn't work for our company."

10 — Section

Closing.

This is not a tool stack.

It's an
operating system.

Three things this week.

Install Claude Code today. Not Friday. Five minutes.
Clone the company repo. Run claude. Ask it "what is this project?"
Show up to the co-work Wednesday. Bring one real question.

Do those three. The rest is downhill.

Want it rolling tomorrow? Book a 15-min intro at cal.com/dima-levin — first slots usually within 48 hours.

I'll see you
in the repo.