Claude CLI Dojo

Four Keys at the Start of a Line

The first character you type tells Claude what kind of input this is.

Heads up on !: it runs non-interactively — no TTY, so sudo password prompts hang. Use passwordless sudo (NOPASSWD) for specific safe commands, or run privileged steps outside the Claude session.

Don't Start Over. Resume.

Every session is saved. You almost never need a fresh start.

Give Your Session a Name

The picker is full of "untitled" sessions. Three seconds to name one saves you ten minutes later.

> /rename auth-bugfix

Sets the name. Shown in the prompt box,
the picker, and your terminal title.

> /rename

Claude reads the chat and suggests a name.
Confirm or edit. Good when you forgot to
name it at the start.

You can also pass -n <name> when you start: claude -n "caching spike".

Color Your Sessions

When five Claudes run in five terminal windows, color is how you tell them apart at a glance. No more pasting code into the wrong repo.

> /color cyan
> /color #FF6699

Sets the color of the prompt box.
Takes color names or hex codes.

> /color

Picks a color for you. Run it again
if you don't like the one you got.

Pick the Right Model and Effort

Defaults are fine. Knowing when to switch saves money and time.

/clear vs /compact

When a session gets long and slow, you have two options.

Teach It Once

Drop the right file, Claude loads it on every session. Stop explaining the same thing every day.

One File, Every AI Tool

If you only write one rules file, write AGENTS.md. It works in Claude, Cursor, Aider, and others. It lives in git and gets reviewed in PRs.

# AGENTS.md

## Testing
- pytest, not unittest. Mock HTTP with respx.
- Integration tests hit a real DB, never a mock.

## Naming
- Files: snake_case.py. Classes: PascalCase.
- Internal CLI tool: `aihub-ci` (not `ci-helper`).

## Tone
- No em dashes. Use periods or middle dots.
- Spell out tool names: "Dependency Track", not "DT".

## Architecture
- See `docs/architecture.md` for the data flow.

Stop Clicking "Yes" 50 Times a Day

By default Claude asks before every change. Press Shift+Tab to cycle modes.

What It Does, and Why It Is Not Real Security

Claude can run shell commands inside a sandbox so they cannot touch the rest of your system. On Linux it uses bubblewrap, on macOS sandbox-exec. It is useful, but read the right column before you trust it.

The Ones You Will Use Every Week

Type / to see all of them. These are the keepers.

Run Two Claudes Without a Mess

A git worktree is the same repo checked out in a second folder on a different branch. claude -w creates one and opens a session inside it — isolated files, shared history.

# Make a new worktree and open Claude
# inside it.
claude -w spike

# Same, but drop it in its own tmux
# window so it has its own keystrokes.
claude -w spike --tmux

# Outside Claude: list and clean up.
git worktree list
git worktree remove spike

Mount .claude Into the Container

A devcontainer is wiped on every rebuild. Without a mount, your sessions, memory, and login die with it. One line in devcontainer.json keeps them.

{
  "mounts": [
    "source=${localEnv:HOME}/.claude,target=/home/vscode/.claude,type=bind"
  ]
}

Claude in Pipes and Scripts

claude -p runs without the chat UI. Use it like jq or curl.

# One-shot question.
claude -p "summarize the last 5 commits"

# Pipe stdin in.
git diff | claude -p "explain this diff"

# JSON output for scripts.
claude -p --output-format json \
  "list TODOs in src/" | jq '.result'

--output-format json
--max-budget-usd 0.50
--model haiku
--allowedTools "Read,Grep"

Warn Me Before I Run Out of Context

Claude shows a status line below the prompt. Hook in a small script and it reports live token usage. Above 200k it tells you to run /compact.

#!/usr/bin/env python3
"""Status line: model · tokens. Warn near the limit."""
import json, os, sys

data = json.load(sys.stdin)
model = data.get("model", {}).get("display_name", "claude")
path = data.get("transcript_path", "")

tokens = 0
if path and os.path.exists(path):
    for line in open(path):
        u = (json.loads(line).get("message") or {}).get("usage") or {}
        if u: tokens = sum(u.get(k, 0) for k in
            ("input_tokens", "cache_read_input_tokens", "cache_creation_input_tokens"))

if   tokens >= 200_000: c, h = "\033[31m", " /compact"
elif tokens >= 150_000: c, h = "\033[33m", " !"
else:                   c, h = "\033[32m", ""
print(f"{model} · {c}{tokens // 1000}k{h}\033[0m", end="")

{
  "statusLine": {
    "type": "command",
    "command": "~/.claude/statusline.py"
  }
}

Expert Sub-Modes for Specific Tasks

A skill is a focused instruction set Claude switches into for one well-defined job. Same model, but loaded with the right context and tools for that task.

Give Claude More Hands

Two ways to teach Claude tricks beyond what ships in the box.

claude mcp add <name> <command> [args...]
claude mcp add-from-claude-desktop   # import all
claude mcp list / get / remove
claude mcp add-json <name> <json>     # complex configs

claude mcp add playwright npx -y @playwright/mcp@latest

> open localhost:3000, log in as alice, screenshot
> click "Sign up", fill with junk, tell me what breaks