Edition 11.0 · Jul 30, 2026The foundations edition. Every other page here tells you how to operate the tools; this one explains what is happening inside them, using the domain people find most magical. Sound is already numbers — 44,100 of them a second — and the whole breakthrough was compressing that into a few hundred learned symbols a second, short enough for a machine built to finish sentences to finish a song instead. Then the interesting part: the same three-step recipe went and conquered proteins, robot motion and the weather, which is the actual event of the last few years. Shipped with an interactive lab that runs the same pipeline across six alphabets, and with the parts most retellings get wrong kept in — including five claims from my own first draft that failed verification.
Edition 10.9 · Jul 27, 2026The refresh edition. Two Anthropic models shipped since the last capture — Sonnet 5 on June 30, Opus 5 on July 24 — and one of them broke the rule this site's tier list was built on: for as long as these leaderboards have existed, the most capable model was also the priciest. Not any more. Opus 5 took the top of the independent index at LOWER cost per task than the model below it. Every board on Ch 24 is re-captured, the vendor cards now show the public leaderboard beside the launch claim, and there is a new two-page model file on the model itself — with the part most coverage skipped: the launch table runs at max effort, the API default is high, and on debugging the expensive setting scores lower.
Edition 10.8 · Jun 23, 2026The SEO reading. Most "agentic SEO" wins are self-reported traffic graphs that crater in six months. This chapter pulls the real receipts from two products built by AI agents end to end: Folderly EmailGen — a brand-new subdomain that hit #1 for "business email generator," 467K impressions, third-party traffic value climbing ~$0 → ~$320/mo — and LinguaLive, the cautionary tale whose 153K-impression graph looks like a win until you see ~94% of the clicks are people already typing the brand name. The method is triangulation: GSC for footprint, Ahrefs for the independently-checkable commercial keywords, conversions for truth. Do not trust a graph you cannot cross-check.
Edition 10.7 · Jun 18, 2026The builder's eval, one altitude above Chapter 25. Chapter 25 was a three-line smoke detector for a skill's artifact; Chapter 47 is for when the model's output is the product itself. The Anthropic API course ships, as working code, the two tools most builders never build — a way to SCORE output and a way to RETRIEVE the right context. Both ported to TypeScript over a weekend: a PromptEvaluator (synthetic dataset → cheap code-grade + LLM-judge → honest scorecard) and a hybrid BM25+vector retriever. The point isn't the 400 lines; it's that you can stop charging for AI output you can't score.
Edition 10.6 · Jun 16, 2026The showcase. Chapter 46 argued that taste is the last mile; /good-taste makes it visceral — a side-by-side of generic-AI slop (purple glow, floating blobs, a fake dashboard) next to the same brief art-directed with a public skill. The skill is Leon Lin's open-source taste-skill (github.com/Leonxlnx/taste-skill, 45k+ stars), not mine; this page is what stealing a skill well, with credit, looks like.
46 chapters · 25 interactive widgets · 3 real embedded case studies · the source is open and is the recipe. Read it, fork it, copy the recipe — no signup.
“It has always given me satisfaction to be able to create what I have in mind.”
— Vlad Podoliako
The thesis
HTML-ization. Stop sending dead files.
Every deliverable — report, pitch, audit, deck, model — ships as a live interactive HTML artifact on a deployed link, not a PDF attachment that started rotting the second it was exported. The link is current because the repo is. People forward links; they archive PDFs unread.
This page is the proof. The book about building this kind of artifact is, itself, this kind of artifact — and the source you're holding is the recipe.
single-file HTML · ~114 KB · client / domains / infra de-identified
A real Folderly-methodology audit of a ~90-domain / ~5K-mailbox cold-email estate — swarmed in days, delivered as a clickable doc instead of a 40-page PDF.
If you scrolled this far, you might be the kind of person who'd appreciate the side project I'm not technically supposed to be working on yet.
An Autonomous Fighting Championship. Humanoid robots, an octagon, blacklight and bass, no humans bleed. The kind of idea you say out loud at dinner and watch the table decide if you're serious. I have a pitch deck. It exists because the deck got built before the next meeting did.