← Work

Ponte

A language that makes AI keep what you decided when it builds for you

Period
Sep 2026 –
Role
Planning, language design, decisions (AI writes the code)
Status
Public

v0.3, MIT license

決めていないことで止まった画面

Tools used

01

Problem

When you hand development to AI, it fills whatever you didn't decide with guesses.

02

Observation

  • In my research, a learning rate I had passed in was silently dropped inside a library.
  • Two days of experiments were wasted before I noticed, and not a single error appeared.
  • In TeaMy too, conditions written in the prompt weren't kept.
  • I compared asking in Japanese with asking in Ponte, six runs each on the same requirements.

There was no way to check that decisions were actually kept. The accident in my research has the same shape as ones in apps and external APIs.

03

Flow

You write data shapes, permissions, and how events are handled, and the language makes sure those promises hold. Anything undecided stops with an error before it runs.

1Write the specWon't run while any tbd is left
2ponte checkChecked before it runs
3The lower layer runsPython for now
4confirmChecked against the values actually used
04

Screens

決めていないことで止まった画面

check result

Stops if even one is left. check won't pass while anything is undecided.

Ponte の仕様から動かしたボードのアプリ

An app run from a spec

Unwritten actions can't happen. No one has a permission that isn't written.

05

Decisions

  • Don't run while anything is undecidedTo stop guesses from filling the gaps.
  • No GUIKeep it a written language, with less to learn.
  • Not built around AIHow people use AI will keep changing.
06

Result & lesson

I had AI build the same requirements several times and compared. Every test passed in 1 of 6 runs when asked in Japanese, and in 6 of 6 runs with Ponte.

TechPython 3.11 (standard library only), GitHub Pages

LinksOpen the site ↗Try it in the browser ↗GitHub ↗

Other work