Vlad - 惯性聚合

A field manual

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

3-10B tokens / mo · 5 companies · 47 chapters · 25 widgets · $0 to read · no email gate

Before

Ten tabs. Twenty minutes. No decisions yet.

gmail.comhubspot.comapp.slack.comnotion.socalendar.google.comsentry.iodashboard.stripe.comapp.ahrefs.comchat.openai.comlinear.app

After

One channel. One coffee. The week is closed.

#ops · 6:30 AM

1.HubSpot — 2 deals advanced overnight, 1 went dark

2.Folderly — deliverability holding 96%, no incidents

3.Mentee A — prep doc generated, three open threads

4.One investor on calendar at 2 PM, conflict resolved

The why

It has always given me satisfaction to be able to create what I have in mind.

— Vlad Podoliako

Forty-Six Chapters

Read straight or jump where you need it. Each chapter stands alone.

Now · Edition 10.7 · June 2026

Vlad's Playbook is public. 47 chapters, 3 live embedded case studies, the source is open and is the recipe.

Read the launch

New · Live · Updated hourly

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.

New here? Start here

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.

New · Model file

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.

Dynamic Workflows

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.

New · Showcase

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.

Memory

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.

New · Method

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.

New · Reference

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.

New · Comparison

Claude Code vs Codex

Ran both in anger. Verdict per job — including the night shift Codex actually wins.

New · Reference

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.

New · Setup

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.

New · Examples

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.

Sovereign Stack

Don't get deprecated

Open-weights LLMs that survive the deprecation calendar — Ollama, the hardware ladder, the heretic question, and the nano-gpt Saturday.

Method

HTML-ization

Stop sending dead files. Every report, pitch, and audit ships as a live interactive artifact — two real ones embedded and clickable.

The launch

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.

Live diary

Launch week

The launch as its own experiment in the thesis. Four surfaces, seven days, receipts accruing — honest, even the embarrassing ones.

Resources

Vault library

CLAUDE.md skeletons, .mcp.json examples, hook scripts, SKILL templates, the five reusable prompts. Copy, paste, ship.

Reference

Cheat sheet

Slash commands, settings, env vars, file paths, keyboard shortcuts. Print it. Tape it next to your monitor.

Vocabulary

Glossary

Fifty-plus terms you'll need, grouped A–Z. Linked inline throughout the book. Hover any underlined term to read it.

Tier list

Rank without mercy

Drag-and-drop your own tier list across AI tools, connectors, and infra. The live LMArena leaderboard for reference.

Orchestration

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

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.

The stack

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.

Research

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.

推荐订阅源

Hacker News - Newest: "AI"