Prompts, skills, loops, routines: the four levels of Claude Code automation

The four levels of automation

This is the organising idea of the course, and the reason to read this page even if marketing is not your use case. Each rung removes more of you from the loop:

1. Prompt — you type, it does the thing. Requires you at your computer.
2. Skill — a domain-specific program that runs on knowledge work. Instead of describing the task each time, you invoke it by name and it supplies a long, detailed prompt for you.
3. Loop — that skill run repeatedly at an interval.
4. Routine — a templated skill kicked off on a schedule, by API or by webhook, running in the cloud rather than on your machine. This is the level where the work happens without you present.

His worked framing: build a dashboard skill that pulls the Google Analytics API, a data spreadsheet and your agency analytics tooling, then schedule it as a routine to populate the dashboard once a day.

Step-by-Step Breakdown

1. Create the account and pay

   Sign up at claude.com. He is direct about the money: Claude is usable free, but Claude Code is not — the code surface is what makes the high-leverage automation possible, so a subscription is required. He quotes his own cost at roughly $24–28 CAD per month. Select "personal use" during onboarding unless you are actually on a team plan.

2. Install the desktop app

   The download button detects your OS. On Chrome the download is often blocked — use "download unverified file" in the top-right corner. On macOS, drag the app to Applications and approve the "downloaded from the internet" prompt on first launch. Sign in again with an emailed verification code.

3. Switch from Home to Code

   The single most important navigation point for anyone new: there are effectively two products in one app. Home is the simpler chat product. Code is what you want. The interface changes substantially when you switch, and the Code tab opens on a usage dashboard — sessions, messages, total tokens, and per-model breakdowns over 7 days, 30 days or lifetime.

4. Understand the permission prompt

   When Claude wants to run a command it asks, with three options: deny, always allow, or allow once — with allow-once highlighted as the default suggestion. His rule of thumb is to reserve always-allow for low-risk operations. The expandable tab during a run shows the actual commands being executed under the hood, plus elapsed time and tokens consumed.

5. Open Customize → Skills

   This is where the leverage is. Anthropic ships example skills — morning, plus setup, writing-style and skill-creator meta-skills. He walks the morning skill to show the shape: a description of exactly when the skill should and should not fire, and a long prompt body that gets injected into Claude when you invoke it with /morning. That is the mechanism — a skill is a named, reusable prompt program with its own trigger conditions.

6. Add connectors

   Connectors hook external apps into Claude — Gmail, Google Drive, Slack in the demo, and Spotify as a live example where the connector exposes discrete functions (get current track, player state, volume, position, repeat, shuffle). Connectors are how a marketing automation reaches your actual data sources.

7. Review plugins

   Plugins are packaged extensions from Anthropic and partners, each with a source, version and author. Examples shown in the directory: finance (journal entries, financial statement reconciliation), legal (contract review, NDA triage, compliance workflows), bio research (pre-clinical databases), and code modernization for legacy code. They range from changing Claude's output style to adding structured actions similar to skills.

8. Check memory

   Memory is the profile of basic identity facts Claude keeps about you, built up from conversation. Worth reviewing so you know what is being carried between sessions.

9. Pick your execution environment

   Four options: local (your machine), cloud (Anthropic's), remote control (another computer you own), and SSH. Local and cloud are the two that matter for most people.

10. Store credentials as environment variables

    Each environment has a settings gear where you set environment variables — written in the intimidating NODE_ENV=production style, but functionally just stored credentials. This is where API keys go for services your automations call, such as the Meta Ads API. Values are masked back to you after saving.

Local vs cloud environments — the actual argument

The reason he gives for cloud environments is the one that generalises: your machine is configured differently from everyone else's, so something built locally may simply not run for a colleague. A cloud environment standardises it — same server, same environment variables, same setup scripts — so the thing behaves identically regardless of whose account runs it. He frames this as mattering mainly for larger teams, and expects most solo work to stay local.

Artifacts as a hosting shortcut

A practical trick for anyone doing client work. Artifacts are self-contained web pages Claude builds; once created, an artifact can be opened in your browser and shared by link — either to your team on a Teams/Enterprise plan or to anyone on the internet. He demonstrates the shared link loading in an incognito window with no sign-in. His use for it: he builds internal tools and web dashboards for clients, and artifacts remove the need to host a site at all.

Gotchas & Caveats

• The interface is a moving target. He is explicit that Anthropic tests a lot of UI features and that many are short-lived, so buttons will not match what you see. Learn the concepts — Home vs Code, skills, connectors, plugins, environments — rather than the button positions.
• Tokens are the budget you are actually managing. Bigger projects consume more, and consumption is what determines both your bill and your remaining usage. The in-app dashboard exists so you can see it; use it.
• Autocomplete lets Claude drive your side of the conversation. Pressing tab accepts a suggested continuation of your own message. He estimates it guesses his intent 30–40% of the time and notes plainly that accepting it is not always the right move.
• Generated content can be confidently wrong about you. The demo artifact — a bio page built from publicly available information — got his education history wrong, apparently from a popular but inaccurate post. A useful reminder for anyone generating client-facing material from web research.
• Environment variables are credentials. They are stored per-environment, and the setup walkthrough puts real API keys there. Treat that store with the same care as any secret manager, and see our security centre before wiring production credentials into an agent.

Key Takeaways

• The automation ladder — prompt, skill, loop, routine — is the mental model to carry into any agent platform, not just Claude Code.
• A skill is a named prompt program with explicit trigger conditions; a routine is that skill on a schedule, API call or webhook, running in the cloud without you.
• New users almost always start in the wrong place: switch from the Home tab to the Code tab.
• Customize is where the real configuration lives — skills, connectors, plugins and memory are all one click deep.
• Cloud environments exist to make a build reproducible across machines; environment variables are where API credentials go.
• Artifacts can be shared by public link, which removes the hosting step for client dashboards and internal tools.