Launch an interactive session in any repo.

Launch with the brief straight in.

Continue your most recent conversation.

Resume — pick any past session from a list.

Interrupt mid-task and redirect.

Rewind to an earlier point, take a different path.

Cycle modes: normal → auto-accept → plan mode.

It proposes before it builds. Approve, then release.

Trigger deeper reasoning on hard decisions.

Maximum reasoning for the genuinely hard calls.

"Rate-limit the API: 100 req/min, 429" beats "make it robust".

"Done = tests pass, page loads under 1s."

Drag images into the terminal — "make it look like this".

Generate CLAUDE.md — the brief read at every launch.

Standing orders: stack, conventions, red lines.

Global memory — personal rules across every project.

Start a message with # to save to memory.

Wipe context between unrelated tasks.

Compress a long session, keep the essentials.

Point at exact files instead of letting it search.

Session spend in real time.

Cheap models for grunt work, big guns for architecture.

Account, model and session state at a glance.

Control what Claude may do without asking.

Health-check the install when something feels off.

Reusable playbooks in .claude/skills — SOPs, executable.

Connect Slack, Drive, databases, browsers, your CRM.

Custom slash commands — /deploy, /report, /brief.

Auto-run lint, tests or notifications on every action.

Define subagents — researcher, reviewer, tester.

Have it review a pull request before you merge.

Parallel sessions on separate branches — no collisions.

Headless mode — pipe Claude into scripts and pipelines.

Claude in GitHub Actions: auto-review, auto-fix, triage.

claude -p "summarise overnight commits and open issues"

Always. It's your name on the commit.

Increments are cheap to roll back; big bangs aren't.

Use env vars. Claude reads your terminal.

"Write the tests, then make them pass."

--dangerously-skip-permissions only inside a container.