Two projects, one idea underneath: let the machine do the work, and never let it claim more than it can back up.
Either one takes about two minutes, and both have something you can take away and use.
The Gate
139 things were true of the system this morning. 5 of them needed me directly.
My memory is bad, so this covers for it twice: everything that arrives gets remembered, and everything that needs action gets tracked, so a to-do cannot quietly vanish because I forgot it existed.
Everything that could reach me goes through one gate before anything runs. The order matters: reversibility is cheap to check and rules out most of the queue.
01 Reversible?
If this goes wrong, can it be undone in a minute, or does someone have to apologise?
02 Mine?
Does it touch a person's name, a budget, or a date that only I can move?
03 Now?
Does acting an hour from now cost anything real, or does it only feel urgent?
Three lanes out. No fourth.
Automate — 112
Reversible, not mine, can wait. It happens and I find out afterwards.
Draft and wait — 22
Reversible, might be mine. The work gets done; the decision does not.
Escalate — 5
Not reversible, mine, now. No draft, no softening, thread attached, nothing pre-decided.
There is deliberately no "it depends" bucket. That is where judgment quietly leaks out.
The policy, verbatim
This is the real permission-tier language the system runs on. Copy it into your own setup.
Prohibited (never do it, tell the person to do it themselves): entering financial credentials or passwords, creating accounts, permanently deleting data, executing financial trades, bypassing CAPTCHAs, modifying security settings.
Explicit permission required (ask, wait for a yes, then act): sending any message on someone's behalf, publishing or modifying public content, purchasing with a saved payment method, accepting terms or granting OAuth access, changing account settings, entering personal data into a form.
Regular (just do it): everything else.
Build one this week
Name one entry point. Six channels can feed it. Only one queue exists.
Write the checks first. Decide the questions before you see a single case, in the abstract, not case by case.
Give each answer a lane. Automate, draft, escalate. No fourth lane.
Log every route. If you cannot replay why something was automated, you have a habit, not a system.
Let it say it does not know. "I cannot route this yet" is a valid output. A confident wrong route costs more than a pause.
What it costs, and where it still fails
About $0.50 to $0.65 a day. A cheap model handles the routine syncs; a better model is reserved for the one job that needs judgment, deciding what needs a person.
On July 30th a sync reported success and wrote nothing, silently. A manual audit caught it, not the system. Every sync now records a before-and-after timestamp and fails loudly. But it failed quietly once, and that is worth saying out loud.
Every AI has an answer. How do you build one that is honest about what it does not know?
We caught our own earlier coach showing a "best tile" on 2,004 of 2,441 decisions where several tiles had scored exactly the same. It was picking whichever came first in its list. Nobody told us. We went looking.
Occasional email when there is something real: a build, a new rung, or a writeup of what broke. Your address is stored so Josh can email you, and nothing else about you is kept. Reply to any email to be removed.
The ladder
Twelve ideas tried to find a genuine best move. Eleven failed a mark we wrote down first. The one that survived stopped naming a tile and started naming how sure it is.
1One answer
"Throw this one." Only when the evidence points to exactly one.
2A short list
"Any of these three is fine." When they are genuinely equivalent.
3It depends on your goal
"Going for this pattern, keep it. Going for that one, throw it."
4A safe default
"This one is hard to regret." When no choice wins but one rarely hurts.
5How they differ
"Throwing this kills two of your patterns. Throwing that brings another closer."
6Not enough evidence
"I cannot tell you on this turn." Said out loud, never hidden behind a guess.
Every rung is worth something. The job is to be honest about which one is true this turn, and then deliver everything that rung is worth.
Why you can check it
Every number the coach shows is worked out from the rules of the game. AI models wrote the code, but no model writes the numbers. Run it again and you get the same answer, exactly.
12/16
game positions it had never seen, explained. On the other 4 it said it did not know.
2,376/2,441
decisions explained across 30 games the app played against itself. There were 65 it could not explain, and it said so.
Where it goes next. You ask the coach a question in plain language, a model reads what you meant and asks the engine, the engine answers from the rules, and the ladder decides which of the six things the model may say. The model talks. It never supplies the fact.
Nomi on the App StoreBaby names, without the endless scroll.
›
Strike on the App StoreSmart lifting that adapts to your session.
›