Power-user patterns for Claude Code: CLAUDE.md, skills, and self-improving agents

The piece argues that getting real leverage from Claude Code requires treating it as a programmable agent rather than a smarter autocomplete. The core discipline is giving the model a way to verify its own work — a feedback loop the Anthropic team credits with a 2-3x quality jump. Practical habits follow: use plan mode (Shift+Tab twice) for anything multi-file, have a fresh second session review the first’s plan to dodge context bias, reference files with @path instead of describing them, and brief the model like a delegated engineer rather than pair-programming line by line.

Much of the guide centers on the .claude/ directory as a layered config system. Project scope ships in the repo; global scope in ~/.claude/ describes the user. CLAUDE.md cascades through monorepo subdirectories, while path-gated rules under .claude/rules/ keep folder-specific guidance from bloating every session. Boris Cherny’s own team’s CLAUDE.md is reportedly minimal — build commands, exact run order, single-test invocations, pre-PR steps — with no style lectures. The recommended habit: when Claude errs, tell it to update CLAUDE.md so it doesn’t repeat the mistake. Cherny calls this ‘Compounding Engineering,’ and CLAUDE.local.md (gitignored) extends the same idea to personal PR-review feedback.

Skills are pitched as the unit of reusable expertise: folders under .claude/skills// containing a SKILL.md whose frontmatter is loaded at session start (~100 tokens) while the body loads on demand. Skills can bundle templates, scripts, and helper docs, run inline shell with ! prefixes, and gate themselves with disable-model-invocation for side-effectful commands like /ship. The author recommends new work go into skills/ rather than the older commands/ directory.