Launch Day, And An Edit That Produced Four Rules

The studio started selling. Three things went out: the first Substack issue, a note to my own list, and three posts on X.

What the day cost

345 minutes of my time, measured from session timestamps rather than estimated. Two blocks, 08:48 to 12:44 and 14:47 to 16:36, with the gap in between at the gym.

That puts the week at 11.33 hours against a five-hour ceiling, on a Wednesday. Launch weeks are not typical weeks. Whether the ceiling holds is answered by the weeks after it, and the figure goes on the numbers page either way.

The prices

The subscription is €50 a year, with a monthly option and a founding tier at €150 and €300. The blend works out at €12.20 a year per paying subscriber once the mix and the fees are taken out.

Free readers get the record: every hour, every euro, every decision and the reasoning behind each call. Paid subscribers get the methods in full, in the working detail that makes them reproducible.

The thread that got thrown away

A machine wrote a ten-post launch thread for X. Every post was under the character limit, every figure traced to a file, every claim checked against a rule. I read it and did not like it.

What was wrong with it was that it opened with the studio. The first post led with seven defects found on the first night, which is interesting to me and to nobody scrolling past it.

I dictated a replacement in about ninety seconds. It opens with the reader:

If you are interested in AI and how it can be used productively, if you always wanted to make your own video game, or if you want to see how far I can push this technology:

Three posts instead of ten. The second one tells my existing followers what that account stays about, which is families, parenting, marriage and relationships, so nobody feels bait-switched. The third routes people to the website and the paid report.

The ten-post version is kept under discarded/. The words were wrong. The machinery that counted characters and checked figures was not.

The disclosure that would have been a lie

Every X post carries an AI disclosure, and the checker enforces it. It blocked all three of the new posts.

The forms it accepted were "AI-generated", "AI-written" and "AI-operated". I wrote those posts. Passing the check would have meant claiming a machine wrote my own sentences, which is a false statement in the generous direction and still a false statement.

The rule says AI-generated content is disclosed. It was being enforced against every post on the surface regardless of who wrote it. That is a defect in the enforcement, not in the rule.

So the checker now takes an authorship assertion. Three things hold it closed: the default is still "machine", so forgetting the flag can only over-disclose; it waives that one rule and nothing else, with a test that pushes a personal-life leak through it and requires a block; and the assertion is recorded rather than verified, because nothing can tell who typed a sentence.

Machine-written posts still disclose. They always did.

What my own edit taught

I rewrote parts of the Substack announcement before approving it. Ten differences between the draft and what I sent, and four of them became rules.

The disclosure had been put first. I moved it to a PS. A reader who gets one paragraph should get the argument, not the housekeeping.

"It starts at numbers small enough to be embarrassing" became "We are starting small and simple." Telling a reader the numbers are embarrassing instructs him to be embarrassed by them.

I also added an ask that was missing. The draft named what the studio runs on and stopped; every place the copy says what something costs is a place a reader may want to help.

Fifteen passages from that edit are now kept as a library, so the next piece copies the sentence I approved instead of writing the same idea again, worse.

The empty list

The studio's Substack has no subscribers. Publishing there reaches nobody.

That is why the note to my own list exists, and why the X posts point at the site rather than assuming anyone arrives on their own. The archive being empty is the acquisition problem, and it is the reason the backfill program started that night.

What ran overnight

Night 003 started at 22:22 and finished at 22:36. Fourteen minutes, committed and pushed on its own, against a three-hour budget of which it used eight percent.

It scoped twenty backfill articles and wrote the first one. It also reported five findings against itself, two of which were defects in the instructions I had approved. One of them: four tests around the numbers page had been red since that afternoon, broken by two commits that never touched code and only changed the ledger.

It did not fix them, and it said why: editing a failing test until it passes is the move the QA Director seat exists to catch.

Noah Revoy CEO, Lone Hand Studio