An automation
Jobs that run without you need a clear brief, the fewest permissions, and someone reading the first runs.
Scripts, crons, bots and anything that should run without you. Paste the prompt below into Claude Code in your project folder: it sets up version history, keeps your keys out of commits, writes a short guide for Claude, adds a check it runs before saying anything is done, and installs the foundation and this kit for this project only.
Paste this into Claude Code
You are setting up this folder for an automation, for someone who is new to building with Claude Code. Work through the steps in order. Explain each one in a sentence of plain English as you go, with no jargon. Ask before anything that installs software or changes files outside this folder.
1. **Check this computer.** Run `git --version`. If git is missing, tell me how to install it and wait. This kit also needs: Node.js, for the Playwright browser tool. Check for each, and if one is missing, tell me in plain words what will not work until it is installed and where to get it, then carry on with the rest.
2. **Understand the project.** Ask me my first name. Then, if the folder is empty, ask in one question what I want to build and who it is for, and wait for my answer. If there is code, read it and tell me in two sentences what you found.
3. **Version history.** If this is not a git repository, run `git init` and explain that it lets us undo any change. If git has no name and email set (check with `git config user.email`), ask me for them and set them for this project only with `git config user.name` and `git config user.email`.
4. **Keep secrets out.** Add a `.gitignore` that covers `.env` and `.env.*` (but keeps `.env.example`), `node_modules/` and build output (`.venv/`, `__pycache__/`, `output/`). Create `.claude/settings.json` with permissions that deny reading them: `"deny": ["Read(./.env)", "Read(./.env.*)", "Read(./**/.env)", "Read(./**/.env.*)"]`, so keys never reach a prompt or a commit.
5. **Write `CLAUDE.md`.** Keep it under 40 lines: what the project is and who it is for, how to run or build it, the one command that checks it works (below), and any rule you would otherwise get wrong. Leave out anything you could work out from the code. Put this line first: "Call me <my first name> in every reply." If you ever stop doing that, the conversation has grown too long to keep track of, and I should start a new one.
6. **A check you can run.** A dry run: one command that runs the automation once on sample data without sending, deleting or changing anything real. Later, once it works, schedule it to start by reporting, not acting (see the first week). If there is no code yet, write the planned check in `CLAUDE.md` marked "set up when the first code lands", and set it up then. Once it exists, allow it in `.claude/settings.json` (for example `"allow": ["Bash(npm run build)"]`) and run it before saying anything is done.
7. **Install the tools**, for this project only. First show me the list below and ask for one yes. Then run each line on its own (a line can chain a few commands), one after another, waiting for each to finish, never in parallel. Skills copied in with git are pinned to the version that was safety-checked; plugins install their current version. Claude Code may ask you to approve each one; if it refuses a line, tell me and show me the line so I can run it myself in a terminal.
The foundation, for every project:
- `claude plugin marketplace add https://github.com/multica-ai/andrej-karpathy-skills.git --scope project && claude plugin install andrej-karpathy-skills@karpathy-skills --scope project`: Coding habits from Andrej Karpathy: think before coding, keep changes small and simple, check your work.
- `claude plugin marketplace add https://github.com/anthropics/claude-plugins-official.git --scope project && claude plugin install claude-md-management@claude-plugins-official --scope project`: Keeps `CLAUDE.md` short and current, and saves what Claude learns each session.
This kit:
- `d=$(mktemp -d) && git -C "$d" init -q && git -C "$d" fetch -q --depth 1 https://github.com/anthropics/skills.git 683bc88e56f3e09ba94f7055977f3d3aa499f202 && git -C "$d" checkout -q FETCH_HEAD && mkdir -p .claude/skills && cp -R "$d/skills/skill-creator" .claude/skills/skill-creator && rm -rf "$d"`: Turns a task you keep repeating into a skill Claude runs the same way every time.
- `claude plugin marketplace add https://github.com/anthropics/claude-plugins-official.git --scope project && claude plugin install playwright@claude-plugins-official --scope project`: Gives Claude a browser for web chores and checks. Needs Node.js.
If a `claude plugin` line says the `claude` command is not found (usual in the Claude desktop app), don't run the other `claude plugin` lines or hunt for the command, but still run every other line. Instead, with the same yes, add these entries to `.claude/settings.json`, keeping everything already there and leaving out any plugin we agreed to skip:
```json
{
"enabledPlugins": {
"andrej-karpathy-skills@karpathy-skills": true,
"claude-md-management@claude-plugins-official": true,
"playwright@claude-plugins-official": true
},
"extraKnownMarketplaces": {
"karpathy-skills": {
"source": {
"source": "git",
"url": "https://github.com/multica-ai/andrej-karpathy-skills.git"
}
},
"claude-plugins-official": {
"source": {
"source": "git",
"url": "https://github.com/anthropics/claude-plugins-official.git"
}
}
}
}
```
Then, in step 9, tell me which plugins are switched on but not installed yet, and that I install each from the **+** button → **Plugins** in a new session in this folder (or run its line in a terminal where Claude Code is installed).
If a line fails for any other reason, say why in one line and carry on with the rest. Never install anything that is not on this list. Plugins switch on in the next session: tell me to start a new one in this folder (in a terminal, restart Claude Code), trust the folder if asked, install any that are switched on but not installed yet, and check they are all listed under `/plugin` or the **+** button → **Plugins**.
8. **Commit** everything with the message "Set up with the Vishwaclaudema starter kit". Tell me the copied skills are included, so cloud sessions get them too. Never commit a `.env` file or anything that looks like a key.
9. **Hand over.** In at most ten lines, tell me what you set up and why it matters, including why you will call me by name, and how to finish installing any plugin that is switched on but not installed yet. Then give me this first-week plan, in my own project's terms:
- Day 1: write down the task you repeat most, step by step, as you do it today.
- Day 2: ask Claude to automate it as a dry run on sample data; check every step's output.
- Day 3: turn it into a skill so it runs the same way every time.
- Day 4: schedule it. If it needs files on your computer, use a scheduled task in the Claude desktop app; if everything is online, use a cloud routine (`/schedule`) and add its keys to the routine first. Either way, start with a run that only reports what it would do.
- Day 5: after a few good reports, let the routine act, and keep a log you read weekly.
Finish with: "Tell me what you want to build first, and I'll plan it with you before writing any code."
How it was tested
A fresh Claude agent pasted the prompt into an empty folder and followed it to the letter, as in the Claude desktop app, where the claude command isn't available, on a computer without Node.js. Without the claude command, plugins are switched on in settings for you to install from + → Plugins in a new session (that last click wasn't tested); the pinned skill-creator was copied in and all three plugins switched on in settings.
Did this prompt work? It worked · I got stuck Opens a public GitHub issue; never paste keys or personal details.
Tested 5 Oct 2026: 4/4 installs worked on a fresh Claude Code · adds about 300 tokens to every sessionWhat the prompt installs, and why
- skill
skill-creator ↗
Turns a task you keep repeating into a skill Claude runs the same way every time.
399.9k installs · by AnthropicSafety-scanned 5 Oct 2026 at 683bc88 · reviewed: safe with caveatsTuning a skill's description runs Claude dozens of times in the background, which uses up your plan limits, and it closes any program using port 3117.
Works in: Chat · Cowork · Claude Coded=$(mktemp -d) && git -C "$d" init -q && git -C "$d" fetch -q --depth 1 https://github.com/anthropics/skills.git 683bc88e56f3e09ba94f7055977f3d3aa499f202 && git -C "$d" checkout -q FETCH_HEAD && mkdir -p .claude/skills && cp -R "$d/skills/skill-creator" .claude/skills/skill-creator && rm -rf "$d" - plugin
playwright ↗
Gives Claude a browser for web chores and checks. Needs Node.js.
by AnthropicSafety-scanned 5 Oct 2026 at d182ca4 · scan cleanWorks in: Cowork · Claude Codeclaude plugin marketplace add https://github.com/anthropics/claude-plugins-official.git --scope project && claude plugin install playwright@claude-plugins-official --scope project
How you will know it works
A dry run: one command that runs the automation once on sample data without sending, deleting or changing anything real. Later, once it works, schedule it to start by reporting, not acting (see the first week).
Add-on packs
Find more tools →Your first week
- Day 1
write down the task you repeat most, step by step, as you do it today.
- Day 2
ask Claude to automate it as a dry run on sample data; check every step's output.
- Day 3
turn it into a skill so it runs the same way every time.
- Day 4
schedule it. If it needs files on your computer, use a scheduled task in the Claude desktop app; if everything is online, use a cloud routine (
/schedule) and add its keys to the routine first. Either way, start with a run that only reports what it would do. - Day 5
after a few good reports, let the routine act, and keep a log you read weekly.
Pro tools 4 more powerful tools · never installed by the prompt
For when you know what you are doing. Each one was reviewed; read its warning and set what it says before you install it.
- plugin
zapier ↗
Real actions in 8,000 apps: email, sheets, CRMs, chat.
official marketplaceSafety-scanned 5 Oct 2026 at 5360f15 · reviewed: safe with caveatsConnect only apps you would let an AI post to, and never choose "always allow" for its write tool.
Works in: Chat (part) · Cowork · Claude Codeclaude plugin marketplace add https://github.com/anthropics/claude-plugins-official.git --scope project && claude plugin install zapier@claude-plugins-official --scope project - plugin
firecrawl ↗
Turns any website into clean text Claude can read, search and crawl.
official marketplaceSafety-scanned 5 Oct 2026 at b310aa4 · reviewed: safe with caveatsPages, searches and any files you parse go to Firecrawl's servers and cost credits. It installs its CLI unpinned and pre-approves every firecrawl command.
Works in: Chat · Cowork · Claude Codeclaude plugin marketplace add https://github.com/anthropics/claude-plugins-official.git --scope project && claude plugin install firecrawl@claude-plugins-official --scope project - mcp
perplexity ↗
Web research with sources, from Perplexity's API.
Safety-scanned 5 Oct 2026 at c58e4ad · reviewed: safeWhatever Claude asks, which can include your project details, goes to Perplexity and is billed to your key. Pin the version and set a spend cap.
Works in: Cowork · Claude Codeclaude mcp add perplexity --scope project --env PERPLEXITY_API_KEY=your-key -- npx -y @perplexity-ai/mcp-server@1.3.0 - mcp
glif ↗
Generates images, video and audio from Claude.
Safety-scanned 5 Oct 2026 at a12ee6e · reviewed: safe with caveatsYour prompts, plus any files Claude uploads, go to Glif's closed servers, and generating media spends credits.
Works in: Chat · Cowork · Claude Codeclaude mcp add --transport http glif https://glif.app/api/mcp --scope project