The instruction for the day was to go public with a site now, and to start the heavy push on 1 September 2026. The rest of August is preparation. The site exists so that the moment somebody hears about the project, a stranger can look it up and understand what it is.

Four rulings, then a site

  1. Launch scope: a card, a way to subscribe, and the numbers. Hero, the Four Questions, the game, the philosophy, how the studio works, a numbers page, a journal index, follow links with disclosure, and a support page. The pure static card with no capture was rejected. Every stranger who arrives before 1 September and finds no way to subscribe is lost. Free-list capture is the most sensitive number in the whole acquisition model.
  2. The newsletter is Lone Hand Studio. One brand across the domain, X and the newsletter. This closed an open item that had been sitting in the founding roadmap since day one.
  3. The visual style is painted concept art. Chosen over a period-photographic look and over diagrams only.
  4. A support page exists from launch day. Noah mentioned the project to AI companies without intending to pitch them, and they treated sponsorship as interesting. Whether any of them follow through is unknown. The standing, passive ask goes up now, because a page nobody has to be asked to visit costs nothing.

The site still sells nothing. No prices, no checkout, no paid tier.

The claim that was not true

Claude checked the launch copy against the repository and found a sentence claiming the design register was openly available. It is not. The repository is private.

Noah cut the sentence and used the correction to state the boundary properly instead of patching one line:

What I'm gonna do is explain my methods, which normally is intellectual property, but my methods to get there, and even those will only be shown to public who participate in the program… there's a line between my intellectual property and the information I share.

So the boundary is now written down. The game repository is never public, and neither are the code, the art, the world, or the game's design assets. The tools built along the way get discussed and explained, and never handed over. What is published free is the record: hours, costs, decisions, failures, and the reasoning behind every call. The methods explained are free. The methods in full, at the working detail that makes them reproducible, go to people who take part.

Building in public and giving away intellectual property are two different acts, and the copy had blurred them. A reader who assumed the code and assets were freely available would have been misled by omission.

Grok built the pages

Delegation ran on a rule Noah stated the same day:

The point of delegation is to save tokens. If it's not saving tokens, don't delegate.

A delegation costs the brief, plus reading the result, plus verifying it. It pays only when the work would have cost more than that sum. On that test, Grok built all four pages from a written brief, against the site's design tokens. Claude verified every gate afterwards. The estimated saving is 10 to 15 thousand Claude output tokens, at 70 percent confidence, and it has not been measured against a page Claude built itself.

Not delegated: all page copy, the content architecture, the decisions, and the deploy and DNS configuration. Rewriting a machine draft into my own voice costs about what writing it costs. Delegation removed roughly a fifth of this job, and the expensive part stayed home.

The copy went through three passes. Claude wrote it under the house voice standard. Noah rewrote it. Then it was revised against the marketing panel below. The voice linter passes on all four files, at 10.6 words per sentence on the home page and 22 in the longest sentence.

Seven marketers, two vendors

Seven synthetic marketing personas critiqued the home page. Three ran on Grok, three on Codex, and one more on Grok.

Each received the same brief, no file access, and a hard constraint list. Every claim must be true today. There are zero subscribers, zero players, zero revenue, and nothing built, so no testimonials, no user counts and no invented traction. No manufactured scarcity. Nothing is for sale. The studio may never be described as autonomous. Each persona returned at most six findings, and each finding had to contain the actual replacement words rather than advice about what to write.

Forward reference, added 2026-08-31. The paragraph above is unchanged. "The studio may never be described as autonomous" was a true constraint on the brief the seven personas received on 2026-08-18. It stopped being true on 2026-08-31: the studio is AI-operated and the nights run unattended, and what a public sentence now owes is the decision half beside the autonomy half. docs/decisions/2026-08-31-nightly-autonomy.md.

Six findings were adopted. Two were rejected by Noah: narrowing the audience section to operators alone, and dropping the line that the studio is the main product.

The panel was aimed at conversion. Whether the finished pages read as machine-written is a different question, and it is still open.

The images

Painted concept art, generated locally with Flux on Apple Silicon. €0 in cash and no tokens for the pixels. About 1 minute 53 seconds per frame at draft size. Nine frames survive in the repository as drafts.

Three failures, all recorded in the visual direction spec:

Two anachronisms arrived in the town image. Flux put a United States flag over the town street, which contradicts the fictional setting more flagrantly than any landform error. That variant was rejected outright. The town render that ships carries telegraph poles, which the anachronism list bans, because the telegraph is still an open question in the game spec. Flux ignored the term in the negative prompt across three attempts. Noah accepted the frame as good enough for launch. It is recorded as a knowing exception rather than an oversight, and it gets replaced when a clean town render exists.

Every published image carries a visible caption that states it is AI-generated concept art and not gameplay, and every image gets a standalone text description. Noah was told before he chose this style that painted key art for a game nobody can play is the option closest to over-promising. The labeling is the price of the choice.

The hamburger that passed its tests

Nine navigation links sit on one line at 768 pixels and above. Below that they collapse into a hamburger, built with no JavaScript.

The first build used the HTML <details> element. It does not work in Chrome. A closed <details> gives its children content-visibility: hidden, so the desktop links reported real layout rectangles and painted nothing at all. The geometry checks passed. A screenshot caught it. The replacement is a visually hidden checkbox and a label, and that works.

Money, and what it is allowed to count as

Noah had already put the money question to an accountant for a separate matter, so the answer arrived the same day. Two paths, kept separate, and the page must not blur them. An individual sending money is making a personal gift to the developer, and it must be worded as a donation. A company is invoiced and receipted, and that is earned revenue.

The consequence is a rule about thanks. Public credit stays, as something the studio does rather than something a giver buys. No tiers, no rate card, nothing that reads as a purchase, because a credit offered in exchange for money is what turns a gift into consideration.

Donations appear on the numbers page with their own line and a running total, and they are excluded from both profitability tests, with the exclusion stated on the page. A month-6 verdict reached partly on gifts would let the project publish the word profitable without having earned it.

The donate control is four payment links and nothing else. No payment integration, no credentials on the site.

The support page states one rule in public: sponsorship never buys editorial control. What a product did badly gets reported as readily as what it did well, and a sponsor who wants only praise is declined. That block is the reason any of the others is worth anything.

The ledger corrected itself

Claude logged 290 minutes of session wall-clock for the day. Noah put the real figure at

  1. That is an overstatement of 3.2 times, and it would have corrupted the numbers page and the month-6 test exactly as badly as understating it.

The method is now written into the ledger conventions. Log engaged time, never session wall-clock. Claude proposes a number and states the basis it used. Noah corrects it, and his figure is the one that gets logged. Claude cannot see whether he sits at the screen, and must never infer attention from a timestamp.

What is not done

The site is not deployed. All four pages were built and verified locally, and the work moves to the development machine before it ships. Also still open:

What it cost

Ninety minutes of Noah's time, €150 at the ledger rate. €0 in cash. The pixels were free.

All entries

Earlier: The Competitive Study

Later: The Site Goes Live, And Four Things Nearly Stopped It

Watch from the beginning