Hey, it's Cindy 🌱 You commented LEVELS, so here is the whole map: what each level actually is, the exact setup for each one (where to click, what to install, what it costs).
btw the waitlist for AI System Club is now open! If you're keen on using AI to automate your content, video editing, job search, business or build AI agents, feel free to sign up (only if you want to hehe). Ps. the waitlist gets first dibs:)
the map
What the three levels really are
The line between the levels is not "smarter model". It is who decides the next step. Anthropic's engineering team draws the same line: workflows are "systems where LLMs and tools are orchestrated through predefined code paths", while agents are "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks" (Building effective agents, Anthropic).
💬 Level one, the chat model. You ask, it answers, one turn at a time. You are the hands. Every step of the actual work still happens in your browser tabs, your files, your copy and paste. This is where almost everyone lives, and it is genuinely great for thinking, drafting and research.
🛠️ Level two, agents. You describe an outcome and it does the work: opens the files, runs the commands, edits the document, checks the result. You approve, you do not execute. The same request that took you an afternoon of clicking becomes one sentence and a review.
🔁 Level three, agentic systems. The agent runs without you sitting there. It checks its own output against reality (a test that fails, a file that does not exist, a number that does not add up), changes approach, and tries again. You set it up once and it keeps running on a schedule or a trigger.
💡 The one-line version. Level one explains. Level two executes. Level three keeps going after you close the laptop. Moving a task up a level is worth far more than moving to a shinier model at level one.
level one
Set up chat so it stops being a search box
Most people never actually finish level one. They use a blank chat with no context, ask small questions, and conclude AI is mid. Two settings fix that, and both are free.
🗂️ Make a project, not a chat. In Claude, open Projects and create one per recurring area of your life (work, content, study, job hunt). A project holds instructions plus uploaded files, and every chat inside it starts with that context already loaded, so you stop re-explaining yourself.
📌 Write the instructions once. In the project instructions, say who you are, what you are working on, how you want answers (length, tone, whether to push back), and what "done" looks like. Three or four sentences beats a page of vague preferences.
🔌 Add connectors so it can see your real stuff. Go to Customize, then Connectors, and connect the tools you actually live in, for example Google Drive, Google Calendar and Gmail. Connectors are available across Free, Pro, Max, Team and Enterprise, with free accounts limited to one custom connector (connectors, Claude help centre).
🧪 Ask it to interrogate you. The upgrade at this level is not better prompts, it is better questions asked of you. Use the prompt below to make it interview you before it answers.
💬 the thinking-partner prompt
I want to think this through properly with you, not get a tidy answer in one go: [describe the decision or problem in a few lines, including what I have already tried].
Before you give me any recommendation, ask me the five questions whose answers would most change your advice, and wait for my replies. Then give me your actual position, the strongest case against it, and the specific thing that would have to be true for you to change your mind. Where you are working from an assumption rather than something I told you, label it clearly so I can correct it.
📌 Where level one stops. The moment you find yourself copying its answer into another tool, renaming files by hand, or repeating the same request every Monday, the task has outgrown chat. That is your signal to move it up.
level two
Two doors into agents, pick by what you do
There are two real entry points, and you do not need to be technical for the first one.
Door one, Claude Cowork For non-developers. You hand Claude real work in files and connected tools and it completes multi-step tasks start to finish, running in an isolated cloud environment. Available on the paid plans (Pro, Max, Team, Enterprise).
Door two, Claude Code For anything touching code, repos, scripts or your own machine. It reads a whole codebase, edits files, runs commands and handles git. It needs a Pro, Max, Team, Enterprise or Console account, since the free plan does not include it.
Set up Cowork (no terminal involved)
🚪 Open it. Sign in at claude.ai on a paid plan and open Cowork. It runs on web, desktop and mobile, and your sessions and files are saved to your account, so closing the laptop does not kill the work (get started with Cowork).
📁 Give it the raw material. Upload the actual files the task needs, or connect the tool that holds them. An agent with no access to your real inputs is still a chatbot.
🎯 Hand it one full task, not a question. Describe the finished artefact you want back, and let it plan the steps. It breaks complex work into subtasks and runs them for you.
👀 Review before you ship. Read what it actually did, not only the summary it writes at the end. This habit is the whole difference between an agent that helps and an agent that quietly makes things up.
🛠️ your first real agent task
Here are the raw files for [the thing you do every week, for example "our weekly performance report"]: [attach or point at the files]. I want the finished report back, not instructions for making it.
Work through it end to end: read every file, pull out the numbers that matter, build the summary in the same structure as last week's version which I have also attached, and flag anything that looks wrong or missing rather than guessing at it. When you are done, list what you changed or assumed, and tell me which parts you are least confident about so I know where to look first.
Set up Claude Code (one command)
💻 Install it. On macOS, Linux or WSL run curl -fsSL https://claude.ai/install.sh | bash. On Windows PowerShell run irm https://claude.ai/install.ps1 | iex. Homebrew works too: brew install --cask claude-code. Requirements are macOS 13+, Windows 10 1809+ or Ubuntu 20.04+, with 4 GB of RAM (setup docs).
🔑 Log in. Open a terminal in a folder you care about, type claude, and follow the browser prompt. A Pro, Max, Team, Enterprise or Console account is required.
📄 Give it house rules. Add a CLAUDE.md file in that folder with your standards and preferences. It is read at the start of every session, so you set the rules once instead of repeating them.
🖥️ Prefer buttons to a terminal? The desktop app runs the same engine with a visual diff review, and you can also run it in the browser at claude.ai/code (overview docs).
👩💻 first task in Claude Code
Before you change anything, read this project and tell me in plain language what it does, how it is structured, and where the risky or fragile parts are.
Then do this one task: [describe the change, for example "add a script that renames every file in /raw to the date it was created"]. Plan it first and show me the plan, then make the change, then run it on a copy of the data and show me the actual output rather than telling me it worked. If something fails, fix it and tell me what went wrong instead of quietly working around it.
💡 The honest cost gate. Level two is where the free ride ends. Chat and projects are free, but Cowork and Claude Code both sit behind a paid plan (Pro is the entry point, see current pricing). If you are deciding, pick the door that matches the work you actually repeat, and give it one real task before you judge it.
level three
Make it check its own work and keep running
Level three is not a different product, it is a different setup. Two things create it: a feedback loop it can measure itself against, and a trigger that starts it without you. On the first one, Anthropic's guidance is blunt: "it's crucial for the agents to gain 'ground truth' from the environment at each step (such as tool call results or code execution) to assess its progress" (Building effective agents). In practice that means giving it something that can fail, not asking it to be confident.
Step one, build the loop
🔁 the self-checking loop
Do this task, then verify your own work before you show it to me: [the task].
Define the check first, in writing: the specific test, command, count or comparison that proves this is right, and what a failure would look like. Then do the work, run that check, and paste the real result. If the check fails, do not explain it away: change your approach, run it again, and keep going until it passes or until you can tell me precisely which part is blocked and why. Finish with the evidence, not with a summary of your intentions.
📌 Why this matters more than the prompt. A check that cannot fail is decoration. "Looks good" is not a check. "The file has 412 rows and the total matches the source to the cent" is.
Step two, add the trigger
☁️ Routines, for work that must run when your laptop is shut. Routines are saved Claude Code configurations (a prompt, repositories and connectors) that run on Anthropic's cloud on a schedule, on an API call, or on GitHub events. Create them at claude.ai/code/routines, from the desktop app, or by running /schedule in the CLI. They are available on Pro, Max, Team and Enterprise with Claude Code on the web enabled, and they are in research preview (routines docs).
🗓️ Scheduled tasks in Cowork, for everything non-code. Open Scheduled in the left sidebar, choose New task, then either Create with Claude (it interviews you) or Set up manually to pick the name, instructions, approval mode and frequency (hourly, daily, weekdays, weekly). Available on Pro, Max, Team and Enterprise (scheduled tasks help article).
💻 Local scheduled tasks, when it needs your own files. The Claude Code desktop app can run scheduled tasks on your machine instead of the cloud, which is the right choice when the job touches folders that only exist locally (desktop scheduled tasks).
🧵 Split the job across sub-agents when it gets big. Claude Code can spawn multiple agents that work on different parts of a task at once, with a lead agent assigning the work and merging results (sub-agents docs).
🌱 a routine prompt that survives without you
Every weekday morning, review everything that changed in [the source, for example "my project folder" or "this repository"] since your last run, and produce one short brief.
Write it as if I have no memory of yesterday: what changed, what it means, what is now blocked, and the single thing you would do first today. Use only what you can actually verify from the files and tools you have access to, and where something is unclear, say so explicitly instead of filling the gap. Save the brief to [file or destination] and keep the format identical every day so I can compare mornings at a glance.
💡 Scope it before you schedule it. A routine runs without approval prompts, so what it can reach is decided entirely by the repositories, environment and connectors you attach. Remove every connector the job does not need, and a green run status only means the session finished, not that the work was right, so open the first few runs and read them.
be honest
Which level are you actually on?
Not a vibe check. Look at what happened after your last three AI sessions.
Level one You got text back and then did the work yourself. Nothing outside the chat window changed.
Level one and a half You have connectors switched on but still ask one question at a time. The tools are there, the habit is not.
Level two You described an outcome, it touched real files or tools, and your job was reviewing rather than doing.
Level three Something ran while you were asleep, checked itself against a real result, and left you something to review.
🌱 The move this week. Take one task you did three times this month, and push it exactly one level up. Not all of them, one. Level three on a task you have never run manually is how people build automations that quietly produce nonsense.