The build-in-public competitive study was ordered on the 15th. Noah read it on the 16th. It covers the eighteen months from February 2025 to August 2026, and eleven projects inside that window. Older build-in-public diaries were excluded, because they predate the tooling and the platform economics of today.
Grok 4.6 ran the roster search with web access. Every load-bearing claim was then opened at the primary source and confirmed, because a subordinate's output is a claim and not a result. Each row in the study is marked verified, claimed, or unknown. Nothing was estimated to fill a blank.
Two things happened when Noah read it. He corrected an identification that had been guessed wrong, and he ruled on the format of everything this studio will ever sell.
The number that invalidated one of our own assumptions
The business plan assumed that 5 to 10 percent of free subscribers convert to paid.
beehiiv's State of Paid Newsletters 2026, verified at the source: median 0.62 percent, top quartile 2 to 5 percent, top decile 5 to 10 percent, with explicit advice to plan revenue at 1 to 2 percent. Five practitioner sources Grok found separately agreed with that shape. Substack's older 10 to 15 percent guidance is dead.
The plan had written top-decile performance in as a planning number. The 5 to 10 percent band came from Claude at 70 percent confidence, and it did not survive contact with the data. The research file says so in those words.
Against Noah's 24,000 followers, this is what 100 paid subscribers requires:
| Free to paid | Free subscribers needed | Share of the 24,000 |
|---|---|---|
| 0.62% (median) | 16,129 | 67% |
| 1% (planning advice) | 10,000 | 42% |
| 2% | 5,000 | 21% |
| 5% (top quartile ceiling) | 2,000 | 8.3% |
| 10% (top decile) | 1,000 | 4.2% |
The Base case needs top-decile conversion and capture of a twelfth of the entire follower base. The month-3 gate of 40 paid subscribers, pre-committed the day before, is now more likely to fire than not. That moves the one-time product from a fallback to the primary revenue path.
The project Noah had seen
Claude had inferred Cook The Dungeon from Noah's description of a build-in-public project he had watched. The inference was wrong. Noah supplied the real one: Marque & Reprisal, by Garrett Petersen, trading as Iron Pagoda.
Verified at the Steam page, the game site and the itch.io page. A solo developer, and a roguelike sandbox sailing simulator set in 1522. A 163,842-hex map of the entire Earth, with 28 empires. Cities whose supply and demand react to the player and to hundreds of independent ships. Planned for Q4 2026. A playable browser demo exists now.
He is the closest comparable in the whole study, and the study had been written without him. His game is the nearest analogue to Black Powder Republic that anyone is attempting. He is a PhD economist who built an audience in another field over years, which is the shape that worked in this roster. The profile that failed is the game-industry person starting a channel from zero.
Three things taken from his example:
- A playable browser demo is the artifact. It costs nothing per visitor, and it makes the 21-day artifact rule concrete. This plan has no equivalent.
- Shorts rather than long-form devlogs.
- His AI disclosure is narrow, specific, and gives the reason for each use. His Steam and itch.io disclosures differ, and both are correct, because Valve's rewrite of 2026-01-16 dropped AI coding assistants from the disclosure requirement while itch tags more broadly. Disclosure has to be written per surface.
What he does not supply is an outcome. No wishlist count, no channel size, no revenue, and the game is unreleased. He is evidence of an approach and not of a result, and anything adopted from him is adopted on its logic.
Video, settled
Three projects in the roster went YouTube-first from a cold start. Crusader shipped on 2025-11-13 at €2.99 and has zero user reviews nine months later. Dino Card Hunt is still unreleased. Shaun Roselt's own on-screen counter read minus $2,119 in annual recurring revenue at day 54. Published marketing guides price a real long-form video cadence at 5 to 10 hours a week for the marketing alone, against a total cap of five hours for everything.
Those three are all long-form or livestream projects. A Short cut from footage the studio already produces is a different cost class. Noah ruled: weekly video output is Shorts only. No long-form devlogs, no livestreams, no channel-building. The same clip posts natively to X, which is where the buyer is.
The ruling nobody asked for
Noah then settled a question the study never asked. His words:
As far as doing any course at the end that people can buy, making the course video is probably the least efficient. It's much more sensible to have it as a mixture of text and images. The reason is, is you can copy text and images into your LLM if you don't understand something. It's very hard to ask video questions using a video.
The buyer is AI operators and founders, decided the day before. That buyer reads inside a model. A video course forces him out of the tool he works in, and it goes inert at the moment he has a question. Text and images survive being pasted into his own AI, and stay a live input to it.
The reason sets production rules, and without them the benefit is lost. Markdown is the master format, and PDF and EPUB are derived from it. Nothing essential exists only inside an image. Every image gets a standalone text description. Prompts and configurations ship as fenced code blocks rather than screenshots. The weekly report obeys the same rules, because it is the recurring form of the same product.
It is also the cheaper decision, which is unusual for the better one. Video is the largest per-unit cost in course production, and the two-to-four-hour packaging estimate for the methodology product only holds in text and images.
The AI penalty, and what the evidence says about Noah's reading
Game Oracle studied 9,879 commercial Steam games released between January and October 2025, excluding free-to-play titles and identified slop developers, and controlling for developer experience, publisher backing, game type and release month.
Games disclosing AI use took a 52.6 percent decrease in total reviews. For established studios the drop runs 40 to 60 percent. For inexperienced developers there is no measurable impact.
Noah is a first-time game developer. On this evidence the disclosure penalty does not apply to him.
His reading of it, said on reading the finding:
I think the problem is undisclosed AI usage, but also, AI allows people to build real garbage very quickly, and that's what people are objecting to. No one cares if the game's good, if it's used, if it uses AI. What they care about is if it's not good.
The evidence largely supports him, and it is the study author's own leading interpretation. The study states that it cannot separate player backlash from AI use being a symptom of cost-cutting that damages quality. CODEX MORTIS discloses that AI produced the entire code, text, art, sound and music, and sits at 80 percent positive across 69 reviews. Shrine's Legacy is human-made, was review-bombed as fully AI-generated, and recovered once the developer posted process proof. The accusation tracked perceived quality rather than actual AI use.
One caveat is owed. The 52.6 percent effect survived controls for developer experience, publisher backing, game type and release month. If the penalty were purely quality-scapegoating, those controls should have absorbed more of it. Confidence that scapegoating is the dominant mechanism: 75 percent. No study isolates quality-matched AI and non-AI titles, so nothing available today resolves it.
Operationally it points the same way either way: disclose accurately, and compete on whether the game is good.
A note on how the research behaved
Grok found an uncited claim of an 18 percent sales penalty on a secondary site, and refused to use it. That refusal is why the real study got found: the gap it left was searched, and the primary source turned out to say something much stronger and better controlled.
What the study could not find
- Months-to-100-paid for any cold-start build-in-public newsletter in the window. Nobody publishes it. The plan's central unknown stays unknown.
- Wishlist counts for any game in the roster.
- Independently verified revenue for the largest reported success in the roster.
- A solo game-devlog account that built a large audience from zero inside the window.
One absence is itself a finding. Publishing revenue is common among indie hackers. Publishing costed hours at a wage, plus a running profit and loss, was found nowhere in eighteen months. The ledger at €100 an hour, published, is rarer than it was assumed to be.
What it cost
€0 in cash; the research ran on subscriptions already held. Four minutes of Noah's time: two to identify the project, two to rule on video and product format.