I write about software, the things I build after work, and life in Japan. Every post comes with audio, so you can listen on your commute.
- TechAITools
Telling Opus 5 to "always verify" backfires: I checked /insights' suggestions against the official guide
Claude Code's /insights gave me five CLAUDE.md rules. Checked line by line against Anthropic's Opus 5 prompting guide, half made things worse. What I kept, what I dropped, and why.
- TechAIQAEngineering
To make AI-driven development faster, I built a system that lets me trust the AI
A skill that takes a new project from interview to verified, working app: exit conditions for every step, one human approval gate, three layers of verification, and a rule for when to doubt the plan instead of the code.
- TechAIWorkflowEngineering
I automated the whole development workflow with Claude Code skills and subagents, and delivery got 3x faster
How I designed one orchestrator skill and six role-based subagents that take a feature from design to deployment, and how the skill grew through five stages of real use.
- TechProjectsAIData
I built a 2026 World Cup match prediction AI with TiDB and Gemini
Giving a prediction bot episodic memory with TiDB's hybrid search (vector, full-text, and SQL in one database), a web UI to try it, and what 40 real matches showed: how you make it remember changes how often it's right.
- TechAIWorkflowTools
An introduction to Claude Code dynamic workflows: how they work and what to watch out for, from the official sources
What dynamic workflows are, how to start and monitor them, how approvals and permissions work, what they cost, and where they are available, using only Anthropic's blog and docs.
- TechProjectsAIApple
Building a local LLM chat app that talks to your documents, with only Apple Intelligence and SwiftData
A native macOS app that answers questions about your own PDFs and text files with an on-device model: architecture, a keyword-based RAG, and the problems I hit along the way.