DoozleB

Systems developer


I build systems software — emulators, engines, simulation — and drive AI coding agents through it.

Agents resolve around 21% of Python tasks and around 4% once C or C++ is involved. Almost everything written about agentic development covers the easy end. I work at the other one, and publish what breaks.

Currently

FourShades

Game Boy emulator · C++20 · built in public

A Game Boy emulator written with AI agents, spec-first and test-first, measured against the public Game Boy test ROMs. The pass rate is the scoreboard, and every number gets published — up or down.

cpu instructions 499 / 500
test roms 85 / 165

Latest Posts

499 out of 500, on purpose

The CPU passes every SingleStepTests instruction but one. The one it fails, it fails deliberately.

FourShades

Starting the scoreboard at zero

Why I am building a Game Boy emulator in C++ with AI agents, and publishing every number.

FourShades

All posts

About

I write low-level code — the layer underneath the apps, where you own the memory, the timing and the correctness. Then I drive AI coding agents through it, which is where it gets interesting.

21% of Python tasks resolved
4% once C or C++ is involved
worse, same models, same problems

Agents struggle here for concrete reasons: compiler output is written for humans who can skim, builds take minutes, and the model has read far less cycle-accurate C++ than it has React.

Nearly everything written about agentic development covers the easy end. I work at the other one, and publish what breaks — including the parts where it goes badly.

Projects

Game Boy emulator · C++20 · 2026 – present

A Game Boy emulator written with AI coding agents, measured against the public test ROMs.

2D engine · C++20 · SFML 3 · 290 commits

A 2D sandbox engine written from scratch, with a simulation core that runs headless.

How I Work

Spec first

Every feature starts as a written spec, then a plan broken into tasks small enough to review one at a time. The agent implements against that, not against a vague prompt.

Tests are the contract

Write the failing test, watch it fail for the right reason, implement, watch it pass. Agents produce code that looks right and does not work. Reading it will not tell you which.

In public, failures included

Every claim ships with its artefact — the spec, the diff, the commit, the test output. If a number goes down, it gets published going down.

How I work, in full