Blog
AI Memory Grows Weeds: Why Timely Pruning Matters
Last time I covered how to build a memory system for AI. But building is only the start—the longer a project runs, the more memory piles up, and without prun...
How to make your AI actually get you: give it a memory
I keep tripping over the same AI in the same spot—it writes deprecated syntax I've told it not to, forgets rules I've spelled out. It's not dumb; it simply h...
My AI cried 'prompt injection!' — and I believed it. Then it turned out to be a false alarm
Mid-task, my AI stopped and said: 'I need to flag a security warning first'—a command's output supposedly carried an injection that would exfiltrate my usern...
The AI did the work, but I'm the one who's wiped — you're the controller
The last few posts were all about getting the AI to do more, in parallel—saving tokens, tiering models, orchestrating multiple agents, worktrees. This one's ...
Want AI to work in parallel? First give each one its own workspace
Last post was about parallelism—one big task, split and fanned out to several agents at once. But I left a thread hanging: when several hands edit the same c...
Your AI feels slow? Maybe it's not dumb—you're making it work one thing at a time
Last post was about making the AI work cheaper. This one's about making it work faster—a big task, instead of letting the AI grind through one module at a ti...
AI getting dumber the longer you chat? It's not the model—time to take control
The last two posts were about saving tokens before things hit the context. This one's about two things you do mid-session—one decides who does the work (mode...
AI coding getting pricier? I cut my tokens by 82% (with real data)
Last time I said saving tokens is about using your tools right, not cutting docs. Someone asked: how exactly? Here's the hands-on version with real numbers —...
Your docs aren't burning your tokens — your tooling is
People keep asking: with all those PDLC docs, aren't you burning tokens? Having lots of docs and burning tokens are two different things — and if you want to...
PDLC 1.1: two things v1.0 got wrong about artifact shape
v1.0 solved the vertical problem: keeping a feature's stages in order. v1.1 fixes two horizontal problems — artifact shape confusion and a flat feature names...
aitm 1.0: a terminal where the AI is a participant, not the driver
aitm 1.0 is out. 5.3 MB binary, 3–5 ms cold start, six LLM providers, and a four-layer safety model that keeps your hands on the wheel. Here's why the design...
arcade: a browser-based arcade emulator on this blog
There's a browser-based arcade emulator at /arcade/. Supports MAME and dozens of consoles. Drag in your own local game files and play — everything runs insid...
PDLC: turn AI-assisted development from soft conventions into hard contracts
AI says 'I built the feature' but the PRD only lives in the chat transcript. Tests get 'added later'. Across sessions, nobody remembers which stage anything ...
aitm launch: AI inside the terminal
A desktop terminal app with AI built in, for macOS and Windows. The AI can read files, search command history, and run commands on demand — every high-risk a...
Welcome to my tech blog
First post on my personal tech blog — what I plan to write about, and what I hope to get out of it.