Daily use · From install to fully automated

Write a global ~/.claude/CLAUDE.md ── once, applies to every project

Claude reads this on every project startup. Tell it who you are, what languages you write, your code style preferences. Never repeat yourself again. Template:

My name is Alex. I write mostly TypeScript and Python.
Code style: small modules, clear names, avoid unnecessary abstraction.
Don't add libraries I didn't ask for.
Don't add comments unless the logic is genuinely unclear.
Reply style: get straight to the answer, no lengthy preamble.
If you're unsure about something, ask first.

Per-project ./CLAUDE.md ── project-specific context

The global one says "who I am." The project one says "what this project looks like": tech stack, directory conventions, build commands, things to avoid. Claude reads both when it starts up inside the project.

/init ── don't want to write it yourself? One command auto-generates it

Open Claude in a project root and type /init. It scans the codebase and generates a CLAUDE.md (detected stack, directory layout, key commands). Read through and just add your personal preferences on top.

/help ── any time you forget what's available

Lists every slash command in this session — built-ins, plugin-provided, and project-local. Just installed something and forgot the trigger name? Run /help first.

Shift + Tab mode toggle ── 3-state cycle

With the input box focused, press once for Auto-accept edits (file changes auto-approved, Bash still asks). Press twice for Bypass permissions (full auto, asks nothing). Daily driver: stay on Auto-accept — fast and safe.

/context ── see how much context you've used

Shows total tokens used, what's eating them (MCP tools / skills / messages / system), and free space remaining. When Free space drops below 30%, time to /compact or /clear.

/compact vs /clear ── condense or wipe

/compact: summarize the conversation so far, keep the gist, continue the same task.
/clear: zero out, only CLAUDE.md remains.
Rule: same task continuing → compact. Switching to a new task → clear.

/aside ── side question without polluting the main task

You're deep in something complex and need to ask a quick unrelated question (syntax check, doc lookup)? Use /aside. Claude answers, then automatically returns to the main task — no context pollution. 100× cleaner than opening a new window or running /clear.

Esc to interrupt + Esc Esc to edit the previous message

Claude going the wrong way? Hit Esc to stop immediately. Want to fix your last prompt instead of resending? Double-tap Esc and edit it inline. No copy-paste, no restart.

Auto-accept edits ── how to enable + when it's safe

Press Shift+Tab once. File Edit/Write auto-approves; Bash commands (especially destructive ones) still ask. Use it: TDD, refactoring, writing new features — visible, controllable changes. Don't use it: when editing .env, settings.json, or other sensitive config.

Bypass permissions ── use carefully

Press Shift+Tab a second time and nothing asks for confirmation — including rm -rf, git push --force, deleting emails, changing system settings. Only use when: ① working tree is committed, ② you trust the current task, ③ you're not on a production system or important account. Switch back the moment you're done.

/model ── switch models, save up to 3× cost

Three tiers:
Opus 4.8 (latest flagship, May 2026): main coding + complex architecture, agentic coding 64%→69%, ~4× less likely than 4.7 to leave bugs unmentioned.
Sonnet: balanced, cheaper for daily work.
Haiku: 90% capability / ~3× cheaper — use it for batch refactors, doc edits, JSON cleanup, anything mechanical.
Switch mid-session anytime with /model; arrow keys also adjust the effort level. No restart needed.

Opus 4.8 trio ── /fast · /effort · dynamic workflows

/fast on: Fast Mode — 2.5× faster output, same quality, 3× cheaper on Opus 4.8. Keep it on for everyday iteration/debugging.
/effort: tune reasoning depth across six levels low/medium/high/xhigh/max/ultracode. Mechanical → /effort low, hard problems → /effort xhigh.
Dynamic Workflow: the biggest new capability in Opus 4.8 — covered in full in the next step ↓.

Dynamic Workflows + /workflows ── let Claude run dozens to hundreds of agents at once

What it solves: Normally every intermediate step Claude takes piles into the conversation context, and a few dozen steps blow out your window. A dynamic workflow has Claude write an orchestration script that runs a large batch of subagents in the background — the plan and intermediate results live in script variables, so your context holds only the final answer. This is how Opus 4.8 pushes "fully automated" to its limit.

How it differs from plain subagents: with plain subagents Claude decides turn by turn who to spawn and every result lands in context, so scale stays small. A workflow moves the loop, branching, and results into the script — an order of magnitude more agents (dozens to hundreds) — and is resumable within the same session (completed agents return cached results). The script can also have agents adversarially review each other's findings before reporting, giving a more trustworthy result than a single pass.

Three ways to trigger:
① Put the word workflow in a prompt — e.g. "run a workflow to audit every endpoint under src/routes/ for missing auth" — and Claude writes a script on the spot.
② /effort ultracode — once on, Claude plans a workflow for every substantive task (a single request may split into understand → change → verify). It burns more tokens; drop back with /effort high for routine work.
③ Bundled — currently one: /deep-research <question> (multi-angle search + cross-check + cited report).

/workflows is the progress board: it does not start anything — it lets you watch the background runs. Arrow keys to select, Enter to drill into a phase/agent and see what it's doing; p pause/resume, x stop, r restart an agent, s save the run's script as your own /command for reuse.

Hard limits (official): up to 16 agents concurrently, 1,000 total per run (runaway guard); no mid-run input (split into multiple workflows if you need sign-off between stages); requires Claude Code v2.1.154+, and Pro users must enable it in /config first; resume only works within the same session — exit and it restarts fresh.

How an installed skill works ── auto-trigger, with GSAP as the example

The key idea: drop a skill into ~/.claude/skills/<name>/SKILL.md (personal level, every project) and you never import it or remind Claude. Claude invokes it automatically when your words match the trigger terms in its description. Adding/editing/removing a skill takes effect in the current session — no restart. Forgot what's installed? Ask "what skills are available?" or run /skills.

Worked example ── the official GSAP animation skills (7.5k★): one line installs all 8 modules (core API / timeline / ScrollTrigger / plugins / React·Vue·Svelte / performance):

npx skills add https://github.com/greensock/gsap-skills

After install, zero config — just talk to Claude and it taps the right module:
· "add an entrance animation to this hero" → uses gsap-core
· "build a ScrollTrigger parallax" → uses gsap-scrolltrigger
· "write a timeline sequence in React" → uses gsap-react + gsap-timeline
Bonus: after Webflow's acquisition, formerly-paid plugins like SplitText and MorphSVG are now free, and the gsap-plugins module shows you how to use them.

PostToolUse hooks ── auto-run tests/lint after every edit

Configure hooks in ~/.claude/settings.json. After every Edit/Write, Claude auto-runs the command you specify (Prettier, vitest, tsc). Configure once, every project benefits forever. Example:

{
  "hooks": {
    "PostToolUse": [{
      "matcher": "Write|Edit",
      "hooks": [{
        "type": "command",
        "command": "npx prettier --write $CLAUDE_FILE_PATHS"
      }]
    }]
  }
}

@filename · /add-dir ── feed Claude extra content

@src/api/user.ts: precisely @-mention a file in your message; Claude reads it in full immediately.
/add-dir ../shared-libs: add a directory outside the current project to the accessible scope (by default Claude only sees the current dir).

Parallel sub-agents ── let Claude split into 3 tasks running at once

Tell Claude: "Launch 3 sub-agents in parallel: one analyzes the auth module's security, one reviews cache performance, one type-checks the utility functions." Claude uses the Task tool to dispatch all three simultaneously — same wall-clock time, 3× the output. Always run independent work in parallel.

Have Claude create your GitHub repo ── end to end

Prereqs: install GitHub CLI brew install gh, then gh auth login (browser handshake, once). Then tell Claude:

"Initialize this directory as a git repo,
create a same-named public GitHub repo,
write a sensible README and .gitignore,
then push the first commit."

Claude runs the whole sequence: git init → write README/.gitignore → first commit → gh repo create → push. Done in 90 seconds.

Full PR flow ── change → review → commit → open PR

The standard collaboration loop. Have Claude run each step:

1. New branch: "Create a feature/xxx branch and switch to it"
2. Edit code: ride Auto-accept mode
3. Self-review: /review — Claude grades it from an engineer's lens
4. Commit: "Commit these changes with a Conventional Commits message"
5. PR: "Push the branch and run gh pr create with a clear title and body"