imagile
The School Attached to the Lab Is Open Now
The model bake-off got its final standings. The training platform behind it is now free to read — every module, both games, no account required.
Writing
The blog
Notes on AI-assisted development, agentic workflows, and building software that ships. No thought leadership — see why.
Your Default Model Is a Budget Decision You Never Made
Four fixed assignments, every current model, every effort setting, three attempts each — and the grades barely move while the bill moves by a factor of four hundred.
You've Already Stopped Reading Your Permission Prompts
Anthropic's own telemetry shows people approve 93% of permission prompts — that's not caution, it's a reflex, and the fix is a policy you set before you're tired.
Your Agent's Attention Budget Is Already Spent
Context rot isn't a cliff a model falls off — it's a gradient, and the fix is curating what's in the window, not waiting for a collapse.
Autonomy Is a Dial, Not a Ladder
Chatbot, human-in-the-loop, agentic workflow, and full autonomy aren't stages you graduate through — they're positions you choose per task, and the vendors selling autonomy keep telling you not to default to it.
The AI Didn't Do It. One of Three Things Did.
Model, harness, and tools are three different jobs — and the coding agent's terminal access is why it can do things a chatbot never could.
A Spec Is Not a Plan
Research, Plan, Implement, Verify is just the ordinary shape of software work, finally said out loud — and the spec/plan mix-up inside it is costing teams more than they realize.
The Diff Is the Evidence
An agent can change forty files in ninety seconds and describe it in one confident sentence. That sentence is a claim, not a record — the diff is the only account of the work that nobody wrote.
DRY Doesn't Mean What You Think It Means
Hunt and Thomas defined DRY as knowledge — not code. AI generates duplicated knowledge at industrial scale. The discipline hasn't changed; the surface area has.
MCPs Are Eating Your Context Window
MCP servers promised to make your AI agent infinitely capable. They also made it slower, fatter, and easier to compromise. CLIs are quietly winning.
Your AI Agent Is a Blender
Every part of Claude Code maps to a part of a blender. Once you see it, you'll immediately know why your output came out wrong — and who to blame.
The Most Useful Plugin You'll Write Is About Yourself
Building a Claude Code skill that encodes your own voice and engineering philosophy sounds like a neat AI trick. It turns out to be the most clarifying thing you can do with your expertise — because it forces you to name what you actually think.