Category edition · 6 articles
Working with coding agents in earnest: context, tooling, delegation, and the guardrails that hold when the model does not.
From the archive
6 filed
Claude Code in practice · Part 6
A skill that is always in context is a memory file with extra steps. The engineering is in the description field and in everything you keep out of SKILL.md.
Claude Code in practice · Part 5
Delegating does not make the model smarter. It buys a clean context window and charges you visibility, and every good use of a sub-agent follows from that.
Claude Code in practice · Part 4
A CLAUDE.md rule is a request. A hook is shell the harness runs regardless, which makes it the only place to put a rule you cannot afford to have ignored.
Claude Code in practice · Part 3
A server gives Claude something new to do, not something new to know. Which is why adding one never fixes a model that keeps forgetting your conventions.
Claude Code in practice · Part 2
A slash command is a prompt you got tired of retyping. What changes when it becomes a file is that it can be reviewed, shared, and wrong in only one place.
Claude Code in practice · Part 1
Claude Code spends a fixed slice of the window before you type a word. Most sessions that go badly are budget failures, not model failures.