Skillery
Guide — 10 short reads

Guide

Installation, first sessions, prompts that hold and the two safety rules that actually cost money. No account, no check — read it and close the tab.

Setup

Getting the tools onto the machine, once.

Install Claude Code on macOSTerminal, one command, two minutes
  1. Open Terminal (⌘ + Space, type Terminal).
  2. Run curl -fsSL https://claude.ai/install.sh | bash.
  3. Start it with claude and sign in through the browser.
  4. If the shell says command not found, close and reopen the terminal so the PATH refreshes.
Install Claude Code on WindowsPowerShell, not cmd.exe
  1. Open Windows Terminal or PowerShell (Win + X → Terminal).
  2. Run irm https://claude.ai/install.ps1 | iex.
  3. Start it with claude from the project folder.
  4. In cmd.exe the installer will not run — use PowerShell.
Install Node.js the way tools expectnvm on macOS and Linux, official installer on Windows
  1. macOS/Linux: install nvm, then nvm install --lts, so versions can be switched later.
  2. Windows: the official installer from nodejs.org, tick “Add to PATH”.
  3. Check with node -v and npm -v.
  4. Avoid sudo npm install -g: it writes into system folders and later breaks permissions. Use nvm or a user prefix.
Put the project in git before the agent touches itA clean diff is your undo button
  1. git init and commit the current state.
  2. Ask the agent for changes one task at a time.
  3. git diff shows exactly what changed; git checkout -- . reverts everything if a run goes sideways.
  4. Commit after each accepted change, not at the end of the day.

Usage

How to ask so the answer is usable.

First session in a folderPoint at files, not at the whole disk
  1. cd into the project, run claude.
  2. Start with: “read src/app.ts and the README, then tell me how routing works” — a read-only task, no edits.
  3. Only after that ask for a change, one file at a time.
  4. Keep the session short: a long one drifts, and every turn carries the whole history.
Prompt structure that survivesRole, inputs, constraints, format, example
  1. Say who the model is: “senior editor”, “backend engineer”, “QA”. It changes wording, not knowledge.
  2. List inputs as a block: the text, the data, the file names.
  3. State constraints: length, tone, what to avoid, what must stay.
  4. Ask for a format: a table, a JSON object, three bullets. Add one short example if the shape matters.
Put variables in one placeReuse a prompt without rewriting it
  1. Keep the prompt as a file with placeholders: {{product}}, {{audience}}, {{tone}}.
  2. Change only the placeholder values between runs.
  3. Keep the rest byte-identical: providers cache the unchanged prefix, and a stable prefix is cheaper and more consistent.
Let the agent show its workDiff first, apply second, test third
  1. Ask: “show the diff, do not apply yet”.
  2. Read the diff. If it touches files you did not mention, ask why before accepting.
  3. Apply, then run the project's own check — tests, build, lint — not the agent's description of them.
  4. If a check fails, paste the output back instead of describing it.

Safety

The two mistakes that cost real money.

Secrets never go into a promptEnv vars, .gitignore, rotate on leak
  1. Keep keys in environment variables (~/.config/app.env) or the tool's own settings, never in the chat.
  2. Add .env to .gitignore before the first commit.
  3. Pasted a key by accident? Rotate it in the provider console — chats and logs keep copies.
  4. Never let an agent print a key “to check it works”; check with a request instead.
Read the command before Entercurl | bash runs whatever the server sends today
  1. Check the domain in the URL: claude.ai, nodejs.org, the vendor's own docs.
  2. If a page tells you to paste an obfuscated one-liner, decode it first or skip it.
  3. Prefer installers that are documented publicly over commands from a chat or an ad.
  4. A terminal has no undo: the command runs with your user rights, not with the site's.