The Claude Code internals you use daily but never configured

Never lose work to a crash or accidental Ctrl-C again. Master session flags, interrupt recovery, and fork-session for risky experiments.

Layer global, team, project, and personal rules. Use @include, scoped paths frontmatter, and CLAUDE.local.md to give Claude exactly the right context.

Compaction is automatic — but you can steer it. Use /compact with custom instructions and /context to monitor window fullness before hitting the blocking limit.

Claude does not inject all memories — a Sonnet side-query selects up to 5 topic files per turn. Learn exactly how selection works and what it looks at.

The description field is your only selector signal. Write memory entries that reliably surface when you need them and age gracefully over time.

Go beyond basic skills. Use model, effort, paths, context:fork, allowed-tools, and when_to_use to build intelligent auto-triggered skills.

Five config layers, one rule set. Know which file to edit for user defaults, team rules, personal overrides, and security-critical flags.

27 events, 5 hook types, and exit code 2 — the three things you need to know to automate anything Claude Code does.

Use UserPromptSubmit hooks and shell injection to automatically prepend branch, date, PR status, and env state to every prompt.

Use Stop hooks with exit code 2 to prevent Claude from reporting completion until your tests pass, linter is clean, or any custom check succeeds.

Define specialized agents with scoped tools, permission modes, turn limits, and model overrides. Build a TypeScript refactoring specialist in .claude/agents/.

Isolate risky work in a git worktree. Agents with isolation:worktree get their own branch — reviewed before merging, auto-deleted if nothing changed.

Run long tasks without blocking the conversation. Use run_in_background, Ctrl+B mid-execution, and async agents that survive Escape.

Claude becomes a pure dispatcher with 4 tools. Workers do all the implementation. Use it for large parallel tasks that decompose into independent workstreams.

Run named teams of Claude agents with a shared task queue in tmux pane splits. Grant per-worker permission modes and monitor all agents visually.

Five ways to set the model and which wins. The opusplan alias switches automatically. Append [1m] for 1M context. Use CLAUDE_CODE_SUBAGENT_MODEL in CI.

Exact match, prefix, wildcard, tool-wide — build layered permission rules that give your team safe defaults without blocking day-to-day work.

Ask Claude to schedule recurring tasks. Make them durable. Set up autonomous morning CI check-ins and PR reminders that run while you sleep.

Type "ultraplan" anywhere to spawn a remote Opus planning session in your browser. Local terminal returns immediately. Teleport the plan back or run it as a PR.

Voice input, the BUDDY companion, KAIROS channels, and always-on mode. Features your colleagues have never heard of — and how to turn them on.

Hand Claude a repo-scale job. It plans, decomposes, dispatches hundreds of background agents, checkpoints progress, and resumes an interrupted run from saved state. Run /workflows to view runs.

The Faster ↔ Smarter slider. /effort xhigh for hard tasks, low for trivial ones. A direct token-cost lever orthogonal to model — pair low+Haiku, xhigh+Opus 4.8.