Hey Reader,
Three reads this week, and each one turns on the same distinction: a document that describes a standard, and a standard that something can actually be checked against. A prompt half the internet copied without the two parts that make it work. A config field nothing read at runtime. A voice guide every competitor could claim word for word.
The seven components, and the two nobody copied
Matt Shumer's Gauntlet Loop went viral in late July: a builder agent, a critic that never sees the builder's reasoning, and a loop that runs until the candidate beats a real reference head to head. Half the internet copied the prompt text. Pieter Levels ran it and got "a total mess and chaos of unperformant code" plus a $500 bill. The method has seven components, and two of them appear nowhere in the famous prompt. I researched the whole thing, rebuilt it as a playbook, and ran it four times: Snake against a live-measured Google Snake, Doodle Jump against 2009 sources, a brand-new 3D board game from zero, and a fourth run with no pixels in it at all. Every run's quality tracked exactly one thing, and it is not the model.
Read all four runs →
The config field that was documentation until it wasn't
Every WoterClip persona names a model in its config: Haiku for the Orchestrator that routes issues, Opus for the Backend that writes the code. Until v0.3.0, nothing read that field at runtime. A session runs on one model, so whatever model launched the heartbeat did all of it. Start from a Sonnet session and your Opus-configured backend wrote code on Sonnet, while your Haiku-configured triage sorted labels at Sonnet prices. The docs said so plainly, and honest documentation of a no-op is still a no-op. v0.3.0 dispatches each persona to a fresh subagent on the model its config declares, and every heartbeat report now names the model that did the work.
See what changed in v0.3.0 →
Ask five writers for the same voice, get five drafts
Ask a founder what makes their brand different and you get "we're authentic" or "we're customer-first." Ask their team to write a LinkedIn post in that voice and the drafts sound nothing alike. That gap is why AI content still reads as generic after you feed it the style guide: most voice documents describe a tone every competitor on the list could claim without changing a word. Brand DNA is the pattern that generates the guide rather than the guide itself, and it comes apart into four things you can actually write down. This walks through defining yours with or without a tool, and why it is heading toward something your software calls instead of something your team reads.
Define yours →
Reply and tell me which one was the read for you. I read every reply.
– Alex
P.S. – The conversations behind reads like these happen in the WotAI Skool community. 760+ builders, three live calls a week, free to jump in.
Before you spend a weekend and a token budget on this
You have probably already filed this method as either the real thing or expensive theater. Fair either way, but the evidence that settles it is not in anyone's thread. It is in what came out the other end, and you can have that in about three minutes.
So here is mine, running in a browser, no account and nothing to install. Snake holds a 143 ms opening cadence and no longer replays the moves it banked while the tab sat hidden. Doodle Jump behaves the same on a 60 Hz and a 120 Hz display, which it flatly did not before round one. Wotnopoly is the zero-to-one case the research says to avoid, so watch whether the board reads before it impresses – a blind critic forced that order across three commits, away from the showier version that demoed better.
The fourth run left nothing to play, because it rebuilt that same site's crawl surface against Google, Bing and OpenAI's own documentation. That one is in the page source.
If it holds up, you have a method worth pointing at your own backlog on Monday. If it does not, you just saved yourself the weekend.
