VLAD'S
PLAYBOOK
The Ultimate AI Dive Deep
For operators who want to stop juggling tabs and start running AI like an OS. 47 chapters. One operator. Belkins, Folderly, the Newsletter, and a portfolio of others.
by Vlad Podoliako · Edition 10.7 · June 2026
It has always given me satisfaction to be able to create what I have in mind.
Forty-Six Chapters
Read straight or jump where you need it. Each chapter stands alone.
- 01 AI as an Operating System
The Day I Killed My Tabs
- 02 The Five-Tool Stack
Five Tools, Not Fifty
- 03 Why Claude Forgets You
AI Is a Temp Agency, Not a Genius
- 04 Obsidian as Working Memory
The Vault — Where AI Becomes Useful
- 05 What a Skill Is
Recipes the Chef Reads Before Cooking
- 06 Parallel Subagents and Fan-Out
The Swarm
- 07 Scheduled Tasks
Make AI Work While You Sleep
- 08 Chat, Cowork, or Claude Code?
Three Doors to Claude
- 09 Blast Radius and Key Hygiene
Don't Get Owned
- 10 Hosted Agents, Local Models, Frontier
The Wild Stuff
- 11 Build a Skill in 30 Minutes
How to Build a Skill, End to End
- 12 Connectors and MCP
Types, install paths, custom servers
- 13 Claude Code in 10 Minutes
The 10-Minute Quickstart
- 14 Slash Commands and Settings
The Cheat Sheet
- 15 When to Skip Permissions
Permissions, Sandboxes, and Sharp Edges
- 16 Hooks and Custom Subagents
From Autocomplete to Coworker
- 17 25 Operator Tips
Hard-Won Wisdom from Hour 200
- 18 Headless Claude and CI
claude --print in Production
- 19 Shipping a Product in a Saturday
How to Build Products with AI
- 20 tmux, Worktrees, Named Sessions
Running Six Claudes at Once
- 21 Which Mode Right Now?
Interactive, Plan, Auto
- 22 Resume, Replay, Fork
Session Management
- 23 A Saturday Build, Hour by Hour
Vibe Coding, with the Misfires Kept In
- 24 The Tier List
Every Tool Ranked Without Mercy
- 25 Evals — Smoke, Regression, Golden
Evals or Hope, Pick One
- 26 How Do I Get My Team to Adopt?
Getting Twelve People to Use This
- 27 Voice Agents — STT, LLM, TTS
Phone Number to Production
- 28 Six Failures, Six Bills
The Receipts I'd Rather Not Show You
- 29 Why Is My Bill So High?
Token Math, Caching, Batch, Routing
- 30 When to Drop CC for the SDK
Building with the Anthropic SDK Directly
- 31 Six Stages from Idea to Deploy
Ideation, Foundation, Creation, Polishing, Security, Deploy
- 32 Agent Archetypes (Rick Platform)
OpenClaw, NemoClaw, Hermes
- 33 Browser Agents with Playwright
Login, Click, Scrape, Post
- 34 Persona Agents and the Four NEVERs
Writing on Your Behalf Without Becoming a Bot
- 35 Codex or Claude Code — or Both?
Day Shift, Night Shift
- 36 When Do I Outgrow Claude Code?
Beyond CC — CrewAI, LangGraph, SDK
- 37 Context Files — CLAUDE.md, memory, skills
Where Conventions Live, Where They Die
- 38 Run Until Done
Goals, Loops, and the Evaluator That Tells the Agent to Stop
- 39 Skills You Should Steal
A tour of the 1M-skill ecosystem, the 73% that's broken, and the gaps an operator can fill
- 40 Prompting, or the Knob You Probably Shouldn't Tune
Why most prompt engineering content is wrong for operators
- 41 Send the Link, Not the File
Every Deliverable as a Live Artifact
- 42 Codex on a Loop
The Second Opinion, Proof-Checked While You Sleep
- 43 Codex as Saviour
When a Second Prior Deletes 90,000 Lines and Hardens What's Left
- 44 Dreaming — Memory That Curates Itself
The Surfacer That Proposes, Never Writes
- 45 The App Store Without Swift
A Native iOS App, Real Recurring Revenue, and Not One Line of Swift I Wrote
- 46 Designing with AI
A Model Can Generate Any Interface in Seconds. It Still Can't Tell You Which One Is Right.
- 47 The Measurement Layer
When the AI Output Is the Product, a Three-Line Eval Isn't Enough
Radar — what's moving in AI, ranked by lead time
A self-updating index of what's travelling through the AI ecosystem, ranked by where it sits on the gradient — papers to repos to community to analysis. The receipts are on every row. See it before the crowd does.
Learn the basics first
New to AI, Claude, Claude Code, or Cowork? The official free courses in the right order, a five-word primer, and the learning ladder — then the book.
Fable 5 — the withheld model, buyable
Anthropic shipped the model it said it wouldn't — split in two. Pricing, the full benchmark table read honestly, the fallback architecture, and the June 22 clock.
Opus 4.8 — the swarm that runs itself
Claude writes a script that plans, fans out hundreds of parallel subagents, and verifies its own work. What it is, the validator loop, how to turn it on, and when not to.
Good taste, on tap
Stop your AI from generating generic slop. A before/after of Leon Lin's public taste-skill (45k★, credited) — and the moon-base and three-design receipts I built with the discipline.
Dreaming — the loop pointed inward
A local, propose-only twin of Anthropic's Dreaming: it digests your sessions, surfaces candidate memories that each cite a verbatim quote, and re-verifies every one against the raw transcript — and never writes to memory itself.
The Self-Audit
The swarm pointed at the agent's own setup. Five auditors, an adversarial red-team, 41 findings — including two refutations that would have broken the system if executed. Your config rots faster than your code.
What Claude Code actually costs
Every plan and API rate, verified — then the receipts: the real monthly bill, the $1,108→$4,312 cache mistake, and the three levers that cut it.
Claude Code vs Codex
Ran both in anger. Verdict per job — including the night shift Codex actually wins.
Hooks that can't be talked out of
Ten real configs with verdicts — what each caught in production, and the three hook ideas that failed expensively.
Claude Code + MCP
The setup guide grounded in a real roster: the servers worth wiring, what each one costs and earns, and the ones that got cut.
12 agents that actually ran
A receipts gallery: the book swarm, the 4 AM watcher, the $0.40/min voice agent — every example with a cost and a failure on record.
Don't get deprecated
Open-weights LLMs that survive the deprecation calendar — Ollama, the hardware ladder, the heretic question, and the nano-gpt Saturday.
HTML-ization
Stop sending dead files. Every report, pitch, and audit ships as a live interactive artifact — two real ones embedded and clickable.
Edition 10.7 is public
Typewriter hero, animated stat odometer, three live embedded artifacts, all 47 chapters as a click-anywhere mosaic, a post-credit scene.
Launch week
The launch as its own experiment in the thesis. Four surfaces, seven days, receipts accruing — honest, even the embarrassing ones.
Vault library
CLAUDE.md skeletons, .mcp.json examples, hook scripts, SKILL templates, the five reusable prompts. Copy, paste, ship.
Cheat sheet
Slash commands, settings, env vars, file paths, keyboard shortcuts. Print it. Tape it next to your monitor.
Glossary
Fifty-plus terms you'll need, grouped A–Z. Linked inline throughout the book. Hover any underlined term to read it.
Rank without mercy
Drag-and-drop your own tier list across AI tools, connectors, and infra. The live LMArena leaderboard for reference.
Swarms
Multi-agent orchestration — the four-pattern visualizer, ten swarm skills shipped, seven patterns I use, the orchestration prompts to steal. Three real multi-session screenshots.
GitHub for non-developers
Five non-code uses, eight gh commands you actually need, what you can safely ignore. Six worked examples — Vlad's Playbook, AFC, Karpathy, Simon Willison ×2, profile READMEs.
66 skills · 32 agents · 12 plugins
Vlad's full Claude Code setup. Every skill folder, every custom subagent, every plugin — sanitized and shareable. Freshly pruned by the Self-Audit.
What shifted this week
External findings that move what an operator does Monday — model releases, deprecations, benchmark shifts. Dated, sourced, signal-not-receipt discipline.
More from the shelf
Decisions, setups, and reference — the rest of the Playbook.
Build vs Buy
Sequencing wedge, cost ladder, and a 5-question matrix for what to make vs buy.
For your CFO
600 words to defend the spend — Claude as a labor line, not a SaaS line.
The 12-rule CLAUDE.md
Karpathy's template at 11% mistakes; the operator overlay gets it to 3%.
Cowork setup
Vlad's Cowork connectors and scheduled tasks — sanitized and cloneable.
Vault starter
PARA vs 7 other systems, project-as-entity, and a working vault to clone.
Weekend builds
Eight Saturday-ship recipes — plus the one that's a trap.
Starter skills
Six drop-in SKILL.md files to paste into ~/.claude and run today.
30-day plan
A day-by-day roadmap from clean machine to operator in a month.
Sections
Every chapter grouped by theme — the table of contents, by topic.
About
Vlad, the portfolio, and the newsletter behind the Playbook.