The commands worth memorizing

One idea explains most of these: output quality = context quality. Most "Claude got dumb" moments are context problems — noise piling up, or a session that has run too long. These commands are how you manage it.

Session hygiene

Keyboard

Prompt prefixes

Setup & utility

CLAUDE.md & making it behave

CLAUDE.md is a file of standing instructions loaded at the start of every session — the difference between re-explaining your project every day and Claude just knowing how your team works.

Where it lives

How to see and edit it

• It's just a markdown file — open it in your editor like any other file, or ask Claude: "show me the CLAUDE.md and suggest improvements."
• /init — generates a starter CLAUDE.md by analyzing the repo. Best first move in any project that lacks one.
• /memory — opens your memory files for editing right from the session.
• # prefix — type # we deploy with make ship, never npm publish mid-session and it's saved for every future session.

The rest of the steering toolkit

• Plan mode (Shift+Tab) — Claude proposes a plan before touching code. Reviewing a plan is far cheaper than reviewing a 400-line diff. Use it for anything non-trivial.
• Esc early, Esc often — interrupt and steer the moment direction looks wrong. Don't let a bad approach run to completion.
• Demand evidence — "run the tests and show me the output" beats trusting "done." Make it verify.
• Review the diff — ask Claude to review its own changes for bugs before you merge. /security-review for anything security-sensitive.
• Be concrete — @src/auth/token.ts plus a pasted screenshot of the error outperforms three paragraphs of description.

Packaged expertise that loads on demand

A skill is a folder with a SKILL.md — instructions plus a description that tells Claude when to auto-load it. Skills keep specialized knowledge out of CLAUDE.md and only spend context when they're relevant.

Where they live

Invoke one explicitly with /<skill-name>, or just work normally — skills whose description matches the task fire on their own.

Skills worth knowing about

Stop the permission nagging — two ways

By default Claude Code asks before running commands or editing files. Useful at first, tedious once you trust your setup. There's a blunt way and a precise way to loosen it.

The blunt way (how Jody runs it)

claude --dangerously-skip-permissions

Skips all permission prompts — "YOLO mode." The name is honest: Claude can run anything without asking, so reserve it for work you'd be comfortable letting a fast-moving teammate do unsupervised, ideally in a repo that's committed and pushed. Don't pair it with credentials you care about.

The precise way: allowlists

Run /permissions in a session, or edit the settings files directly. Rules you allow never prompt again — everything else still asks.

// .claude/settings.json — checked in, applies to the whole team
{
  "permissions": {
    "allow": [
      "Edit",
      "Bash(npm run *)",
      "Bash(git status)",
      "Bash(git diff *)",
      "Bash(git log *)"
    ],
    "deny": [
      "Bash(rm -rf *)"
    ]
  }
}

Middle ground: permission modes

• Shift+Tab to auto-accept edits — file edits fly, shell commands still ask. The sweet spot for most work.
• claude --permission-mode plan — start a session in plan mode: read-only until you approve a plan.
• /config — everything else: theme, notification style, model default.

The tricks that surprise people

Everything here is optional — and each one is a "wait, it can do that?" moment.

Real images via Codex

Claude Code can't generate photographic images itself — left alone it builds code-based visuals (SVG, CSS, canvas). But it can shell out to OpenAI's Codex CLI and use it as an image tool:

1. Have a ChatGPT account, install the Codex CLI, and log in once (codex login).
2. Then just ask Claude Code: "use codex exec to generate a hero image of X and save it in assets/".

# what Claude runs under the hood — locally-authenticated, no API key juggling
codex exec "generate an image of a lighthouse at dusk, save as assets/hero.png"

Codex uses gpt-image-2, which follows instructions well and produces real images — and they drop straight into your project folder where Claude can immediately wire them into the page.

Headless mode — Claude in scripts and CI

claude -p "fix all the eslint errors and print a summary"

One-shot, non-interactive, scriptable. Pipe things in, use it in cron jobs or CI steps.

Parallel Claudes with git worktrees

git worktree add ../myapp-feature-x feature-x
cd ../myapp-feature-x && claude

Each worktree is an isolated checkout — run a session in each and work two features at once without them stepping on each other.

Custom slash commands

# .claude/commands/fix-issue.md  →  usable as /fix-issue 123 by the whole team
Find GitHub issue $ARGUMENTS, understand it, implement a fix,
run the tests, and prepare a PR description.

Subagents — keep your context clean

Ask Claude to "use a subagent to explore how auth works across the codebase." The grep noise stays in the subagent; only the conclusion comes back to your session. It often does this on its own — /agents lets you define custom ones.

Hooks — automation around every tool call

Defined in settings.json: run your own command before/after Claude's tool calls. Classic example — auto-run prettier after every file edit so formatting is never a review comment again.

MCP — plug in external tools

claude mcp add my-db -- npx @bytebase/dbhub --dsn "postgres://..."

MCP servers give Claude new tools: query databases, drive browsers, read Figma, hit internal APIs. /mcp to manage them.

One API for your whole backend

Mosaic gives every project a full backend through a single REST API — and with the skill installed, Claude Code knows the whole platform. "Create a project, add a users table, deploy this page" just works. This site is hosted on it.

What every project gets

Every site lives at https://{slug}.mosaic.site the moment it deploys. Trial: 3 projects, 7 days. Pro: $10/mo, unlimited projects.

Sign up (2 minutes)

1. Create an account at mosaic.site.
2. Go to Dashboard → Settings and copy your API key — it starts with mk_.

Install the skill

mkdir -p ~/.claude/skills/mosaic
curl -fsSL https://api.mosaic.site/skill.md -o ~/.claude/skills/mosaic/SKILL.md

1. Open ~/.claude/skills/mosaic/SKILL.md and add your API key near the top, e.g. My API key: mk_xxxxx — Claude sends it as the X-API-Key header.
2. Restart Claude Code.
3. Test it: "using the mosaic skill, list my projects."