LLMs predict the next token — so why does autocomplete still stop at the next word? I wanted one that finished the sentence. The moment that worked, the experiment shifted: not 'AI helps you write,' but 'AI is writing, and the keyboard under your fingers is just playing it back.'

• macOS
• CGEvent tap
• Gemini API
• Supabase

Had to look like normal typing — no overlay, no completion to accept, no visible UI. One predicted character per real keystroke, so an observer watching the hands and the screen would see the same cadence.

1. Spring 2024 — predict the rest, not the next

   Started as plain LLM-powered autocomplete: instead of one token, predict to the period. Within a day the more interesting framing arrived — undetectable AI typing under your fingers. Same backend, very different artifact.

2. Day 1 — the event tap

   CGEvent.tapCreate at .cghidEventTap, listening for .keyDown, passing the event back through unmodified. Once it compiled and was granted accessibility, every keystroke in every app — Mail, the browser, the IDE — was routing through a function I'd written. That was the whole leverage point: macOS hands you every keystroke if you ask.

3. Gemini fills the rest

   Buffered the trailing 500 characters of typed text and shipped them to Gemini with a one-line ask: finish this sentence, end on a period. First response came back coherent and topical, ~200 characters long, sitting in memory with nothing to do with it yet.

4. Buffer as typist

   Wired the response into an injectionBuffer. handleKeyEvent checks the buffer first: if it's not empty, removeFirst, set the in-flight CGEvent's unicodeString to that character, and return the modified event. Real keypress in, predicted character out. The user is mashing keys; the screen is producing English.

5. The APUSH demo

   Opened a Google Doc and typed 'Three APUSH examples of American Imperialism are'. The rest — the Spanish-American War, the annexation of Hawaii, the Open Door Policy in China — appeared under fingers that were now hitting random letters. From across the room you couldn't tell.

   

6. Shelved

   Never used it on an assignment, never finished wiring Supabase past a logging stub. The honest version of this is 'undetectable AI typing,' which is exactly what you don't ship. The point was that it worked.

The leverage point wasn't the model — it was that macOS will hand you every keystroke in the system if you ask. Once the event tap is up, the LLM is just a fancy text source feeding a queue, and the OS does the actual sleight of hand. I left this on the shelf because the cleanest version of the idea is also the version you can't responsibly ship: it turning out well is what made it not-a-product.