Posts
The deep versions. Every claim traceable to a repo.
- I built a local AI agent that plays Stardew Valley on its own
A C# SMAPI mod and a two-model Python harness let an LLM run a real farm through OpenRouter, cheaply, with objectives graded against structured game state rather than the model's own account of what it did.
- How I set up Claude Code to stop starting from zero
A global memory vault that gives my coding agent context across sessions, machines, and agents, plus the small config tweaks that keep a live session sharp. The architecture, the tooling I looked at, and the honest numbers.
- I had to apply through a company's MCP server, so I made my résumé one
Most hiring MCPs make a candidate's agent apply or a recruiter's agent screen. I published my résumé as a remote MCP server instead, so a company's agent can query my work, assess fit, and reach out. The build, the security model, and the honest adoption caveat.
- 6GB of VRAM is enough, if you know where the tokens go
A personal AI agent on a 2018 gaming laptop with 6GB of VRAM. The five budget constraints that actually decide a local agent stack, each measured.
- The prompt caching mistake that costs more than no caching at all
A changing timestamp in your prompt prefix gives you a 0% cache hit on OpenAI, so you just save nothing. On Anthropic the same line makes your bill 25% higher than if you had never cached. Measured, with a reproducible benchmark you can play with.
- If you keep swapping models in production, you need a coercion layer
I swapped Claude Sonnet for an open model ~50x cheaper. The gaps I hit are why a coercion layer is non-negotiable once you keep changing models in a production agent.
- Evidence over claims
What this site is for, and the plan.