Designing codebases for AI agents
Clear boundaries, local tests, and discoverable rules reduce the cost of every change.
// blog
Blog
Notes from building software and dealing with AI tools.
What I changed in my prompts for Claude Fable 5
Long-running agents need explicit boundaries, checkpoints, and proof of completion.
How to spend fewer Copilot tokens without starving the model
A compact task contract reduces wasted context, output, and review time.
AI-assisted coding needs smaller loops
Give agent small jobs. Check every result.
You do not need to keep up with every AI release
Use a small evaluation loop to decide which tools deserve a place in your work.
A team process for using GitHub Copilot well
Shared repository rules, evaluation tasks, and review policy turn individual prompting into an engineering process.
The developer identity crisis
When code becomes cheaper to produce, specification, review, and debugging become more valuable.
Context is an engineering input
Give coding agents evidence from the repository, not a longer story in chat.
The reality of working and studying at the same time
A practical system for planning two workloads without pretending every week has unlimited capacity.