low effort, high power software robots
A MOSFET is the part that drives every robot ever built: a tiny voltage on the gate switches a lot of current. Almost no input, all of the output. That's the deal here. You say a thing once, the world moves every time after.
You talk to mosfet. It builds you a bot. You teach that bot in plain words, the way you'd teach anyone, and the lesson sticks. Next time the thing happens, the bot just does it.
That's the product. Not a prompt you babysit. Not an agent you program. A thing that learns what you want, then keeps it.
Nobody wants to program their agents. Nobody wants a part-time job as Prompt Janitor because GPT-Whatever-Just-Shipped is cheaper, smarter, moodier, and now "answer the phone" means a haiku and then silence.
They want something that learns. Like anyone else they bother teaching.
So the phone rings, and your bot sits there.
☎ *ring* ⬡ ...It heard it. It thought about it. It did nothing. Fine. Nobody told it that mattered.
👤 hey, every time the phone rings, answer it ⬡ the ring from a minute ago? got it. ✓ learned · answer_ringsYou corrected it once, out loud. It pinned that to the ring it actually heard.
☎ *ring* ⬡ "Hello?"Reflex. No tokens. No "let me think." No invoice because a telephone made a noise.
The order is the point. You configured nothing up front. The bot had to encounter a ring before the word meant anything, and your correction landed on the thing it perceived, not on a string you typed at it. That's why the rule sticks instead of sitting in a prompt hoping to match someday.
A prompt re-decides every time. A latch doesn't. Once it flips, it stays flipped, and holding it there draws nothing.
A better model can show up tomorrow. Use it for the stuff the bot still has to think about. What you already told it stays told. Bye bye, rewrite-the-prompt-every-release.
Gate voltage is the whole interface. You don't machine the transistor, you don't rewire the board, you put a small signal on one pin and the power does what you meant. The bar is so easy a baby could do it. If you need YAML, a system prompt, or a two-week tune-up, we already failed.
You don't open a project. You don't wire providers. You talk to mosfet.
| 1 | Make a bot. | Ask. Get one. |
| 2 | Poke it. | Ring it. Listen. Break it. |
| 3 | Teach it. | Correct it in your own words. Watch the rule stick. |
| 4 | Tell us when it sucks. | If mosfet ships a dud, say so. We can take it. |
It asks for a number, a key, a voice in the conversation. You hand it over. You do not excavate final.env.bak.reallythisone. Each bot keeps its own mind. It gets cheaper as it gets smarter, because it stops paying rent to remember something you already said.
A learned rule is a DIP switch, not a prompt. You set it once by hand, it holds with no power, and you can see it and flip it back. Same board, same idea: the expensive part runs once, the switch keeps the answer.
Four claims. If one breaks, we broke it:
| Claim | Means |
|---|---|
| Told once, told forever | Swap the model. The lesson survives. |
| Learning makes it cheaper | A learned rule is not a token spend. Costs go down over time, not up. |
| Each bot keeps its own mind | No shared brain, no cross-contamination, no "why does my bot know that." |
| The teaching is the interface | If the answer to "how do I change this" is ever "edit a file," we failed. |
Yes, there's a Python library under here, the same way there's a board under the gate pin. There's a library under your microwave too, and you have never once imported it. Code is still allowed (development, examples), and if that's your idea of a good time, go nuts. It's a door marked staff only, not the front one.
Most people just wanted Siri, since 2011, to take the note, make the call, learn the house rules, and not forget next week because someone shipped a new adjective.
Talk to mosfet.
mosfet