A weekly never-finished project. Serious builds, silly detours, and honest notes on what AI actually helped with.
We tested a desktop build of our web minigame using Tauri, transparent windows, and window position detection. It worked, climbed walls, and taught us where to set firmer rules before the agent gets creative.
Built a local-only MySQL admin tool for a legacy PHP upgrade, using Node.js instead of phpMyAdmin and Apache, with ad hoc SQL, multi-database browsing, and snapshot comparison for migration planning. Useful, slightly stubborn, and much better once the UI stopped trying to write its own autobiography.
We built a tiny Windows helper for Stream Deck buttons using C++, Win32 API calls, registry tweaks, and Codex/GPT-5.5 to switch recording setups fast, with fewer clicks and less faff.
We built Big Black Box, a native Windows desktop utility for masking most of the screen with a resizable black overlay and optional transparent cutouts, packaged as a 117KB portable exe in plain C++. Small tool, oddly useful, and only slightly dramatic.
We built a browser platformer sidekick for holeaf.uk using DOM elements as platforms, simple collision detection and no game engine. It mostly obeys gravity.
Built a small Go CLI and skill file to break project plans into JSON task files, claim the next job automatically, and keep AI sessions focused with less context. A bit less chaos, a bit more progress, which is usually enough.
We built an AI-powered website builder prototype inspired by Geocities, using chat-driven planning, draft review, approval before publish, rollback, auth, and site provisioning. It worked, mostly, after a generous serving of agent handoffs, context rot, and human supervision.
We rewired a local Penpot assistant to call the Codex CLI instead of an OpenAI compatible API, testing a shell based handoff, local auth, and what that changes for design workflows. Early, messy, and more promising than polite.
Built a small homepage for a dev VM to list project folders and trigger VSCode remote deep links. A quick test of a Codex-backed workflow, a simple browser dashboard, and one very real snag with window reuse.
We built a social content planner to test calendar scheduling, draggable overlapping cards, and an infinite canvas for rough campaign planning. The bigger problem was not the UI, but deciding whether a custom tool beat existing posting platforms, which is annoyingly human.