How I Use Every Claude Code Feature (blog.sshh.io)

A longtime Claude Code user published a detailed “brain dump” of how they run the agent end-to-end, turning practical tricks into a reproducible playbook for engineering teams. The writeup centers on a single-source CLAUDE.md at the repo root (the agent’s “constitution”) — short, guardrail-focused guidance (not a manual), avoiding raw file embeds, preferring positive alternatives to “never” rules, and syncing with AGENTS.md. It covers workflows (avoid opaque /compact, prefer /clear + /catchup or a “Document & Clear” restart), token hygiene (200k window with ~20k baseline in a monorepo; run /context), minimal slash commands (/catchup, /pr), session resumption (claude --resume/--continue and ~/.claude/projects logs), and a disciplined hook strategy that blocks only at submit (PreToolUse checking a test-pass sentinel) while using non-blocking hint hooks for feedback. Technically significant guidance includes preferring Claude’s Task/Explore (master-clone architecture) over brittle custom subagents to preserve holistic context, and treating “Skills”/SKILL.md as a formalized scripting layer that lets agents execute CLIs and binaries rather than relying on fragile MCP/tool abstractions. The post also recommends built-in planning mode (or SDK-driven planners) for large feature changes, and emphasizes making CLAUDE.md concise to force better tooling and simpler CLIs. For AI/ML teams building AI-IDE workflows, these patterns improve reliability, auditability, and token efficiency while reducing human orchestration friction.