From Zero to Hero: An AI Progression Path

I work in a terminal window all day. Email, WhatsApp, Telegram, team chat, my WordPress sites over SSH, Cloudflare deploys, the browser, even an iPhone plugged into the Mac: I drive all of it by talking to Claude Code. I say what I want done, the AI does the work or drafts it, and I read the drafts and say "send."

There's a ladder between that and never having opened ChatGPT. Each rung has a sign that tells you you're standing on it, and one move that gets you to the next. You don't need to be a developer to climb it. The top of the ladder looks technical from outside, but what you do there is describe outcomes in plain language and review what comes back.

Daily Driver AI started as an in-person meetup in Cascais, with the first one in September 2026, and it's now a community where people share what's actually working. Members range from people who barely use ChatGPT to people building with agents. Find your level below, do the one move, and come back when it feels ordinary.

Level 0: Not Using AI Yet

What it looks like: you've heard about ChatGPT and Claude, maybe opened one once, and went back to how you did things before. Questions go to a search engine, drafts start from a blank page, and AI stays something other people talk about.

Sign you're here: you can't name a task from last week that an AI did for you.

The move: pick one real task this week and do it in ChatGPT or Claude. Real means something already on your list, such as an awkward email to rewrite, a long PDF you were never going to finish, or rough notes that need to become an agenda. A made-up test question teaches you little, because a real task has a result you can judge. Describe it the way you'd explain it to a colleague, then read what comes back and decide whether it saved you time. Skip the party tricks and the poems about your cat.

Level 1: Asking It Questions

What it looks like: you use the AI the way you'd use a search engine. One line in, one answer out, tab closed. Sometimes the answer is good and sometimes it's generic, and you can't tell why.

Sign you're here: your prompts are one line long, and you start from zero every time.

The generic answers are the clue. The AI knows nothing about your document, your client, your tone or your goal, so it answers for an average person. You'd get the same from a stranger at a bus stop.

The move: give it context and real work. Paste in the document, the email thread, the spreadsheet export, and say what you need done with it and who it's for. "Here's the thread with our supplier, here's what I want to happen, draft a reply that gets us there" produces something you can send after a light edit. The same request without the thread produces filler. This is the cheapest improvement on the whole ladder.

Level 2: Daily Chat With Context

What it looks like: the AI is open most of the day. You work in projects, so a piece of work keeps its files and instructions together, and you've written custom instructions that tell it how you like answers (short, plain, no preamble). You upload files, talk to it by voice when typing is slow, and use it for writing, analysis and planning. You ask it to critique its own draft and you ask for a second version.

Sign you're here: you'd notice if it disappeared for a day, and you've stopped re-explaining who you are at the start of each chat.

You get a lot out of this level, and it has a ceiling. The AI produces text and you carry it to wherever it needs to go. You copy the draft into your email and the numbers into the sheet. You're the courier between the AI and everything else.

The move: let it act as well as answer. That means connecting it to the places your work lives.

Level 3: Connected Assistant

What it looks like: the desktop or web app you already use is wired into your email, calendar and drive through connectors, so it reads what's there instead of waiting for you to paste it in. Ask "what's on my plate Thursday and what should I prepare?" and it checks the calendar and the relevant threads before answering. Some of these apps can also browse for you. In the ChatGPT desktop app and on the web, the browser opens websites, gathers current information and takes action.

Sign you're here: you rarely paste anything in anymore, and you've granted the AI access to at least one of your own accounts.

This level changes what you ask for. Before, you asked for words. Now you ask for outcomes: find the invoice and put the follow-up on the calendar. Start with read access and watch what it does with your data before you let it write anywhere.

The remaining limit is the app itself. You work inside someone else's chat window, with the connectors they offer. It can't work in the folders on your computer or run anything there.

The move: install Claude Code or Codex and point it at a folder on your computer. These agents run on your machine and work on your files directly. We've written walkthroughs for Mac and Windows:

Pick whichever fits the subscription you already pay for. The ladder is the same on both.

Level 4: An Agent on Your Own Files

What it looks like: you open a terminal, start the agent inside a folder, and talk to it in plain language. It reads the files there, writes new ones, and runs commands. A folder of messy invoices becomes a tidy spreadsheet, and a page of requirements for a side project turns into a working one-page site.

Once you're in it, the terminal is a chat box with more access. You're still not a developer. You describe the outcome, and the agent works out the steps and does the typing.

The skill you're learning is review. Codex starts out asking for approval before it reaches beyond your folder. Claude Code now starts in auto mode, which surprises people who expect a question before every change, and the Claude Code setup guide shows how to make it ask first. Read what the agent proposes, approve or push back, and check that the result does what you asked. Build that habit here, where the stakes are small, because every level above this one depends on it.

Sign you're here: you've had the agent make something that exists on your disk and that you use, and you've turned down at least one of its proposals because it wasn't what you meant.

This is also where you meet the plain text file that makes the rest work. Claude Code reads a file called CLAUDE.md at the start of every session, and it also reads AGENTS.md, which Codex uses for the same job (see the docs on CLAUDE.md and memory and AGENTS.md). It holds what the AI should know about you and how you like to work. The first version can be a few lines.

The move: connect it to the rest of your work. So far the agent can reach one folder.

Level 5: The Connected Terminal

What it looks like: the terminal agent reaches your email, calendar, chat apps, browser, hosting and business tools, and it holds what it needs to know about you. A few mechanisms do the work.

MCP servers are the connectors. The Model Context Protocol lets Claude Code connect to external tools and data sources, and in Codex it lets the agent use tools and information from other systems. In Claude Code you add one with a single command, for example claude mcp add --transport http notion https://mcp.notion.com/mcp, and manage them with /mcp. Codex has codex mcp add. See Claude Code MCP and Codex MCP.

Skills hold your repeatable processes. A skill is a set of instructions, and optionally scripts, that the agent loads when the task calls for it or that you invoke by name. Whenever you catch yourself explaining the same procedure a second time, it belongs in a skill. Write the steps for your weekly report once, and from then on you ask for the report. See Claude Code skills and Codex skills.

Memory files hold what the agent should know about you and your work. The CLAUDE.md or AGENTS.md from Level 4 grows here into a record of your preferences, your projects and your decisions. I keep my rules in a file like that, plus memory files, and when the AI makes a mistake I write a new rule so it doesn't repeat it. It's my judgment written down where the AI can read it.

Hooks enforce rules. A hook runs your script or tool call when the agent reaches a set point in its work, such as scanning for secrets or validating output before a turn ends. The model can forget an instruction in a memory file, while a hook runs every time. See Claude Code hooks and Codex hooks.

Plugins bundle these pieces so you install a ready-made set in one step. A Claude Code plugin is a directory of skills, agents, hooks and MCP servers that installs and loads as one unit, from a marketplace, with /plugin (plugins overview). Codex plugins can include skills and MCP servers and share one directory with ChatGPT; they work in the desktop app, and the IDE extension doesn't support them (Codex plugins).

The browser belongs here too. The Claude in Chrome extension lets Claude Code work in your logged-in browser. Codex's browser is available in the desktop app and on the web, and isn't available in the Codex CLI or the IDE extension. If you'd rather work in an editor than a bare terminal, both tools run in VS Code and JetBrains (Claude Code in VS Code, Codex IDE extension).

Sign you're here: you say "find the last thing this client sent, check the project folder for what we agreed, and draft a reply," and the draft comes back without you opening either app.

The risk at this level is trust running ahead of verification. Once the AI can reach your accounts, a confident wrong answer can change something real. Secrets stay out of the chat. Anything the AI tells you about a live system gets checked against the live system. You decide which actions need your approval, and you keep that list tighter than feels necessary.

The move: stop switching windows. Make the terminal the place where work starts.

Level 6: Hero

What it looks like: you work in the terminal all day. Several sessions run in parallel, each on its own job. Subagents do research and builds in their own context and hand back a summary, so your main session stays clear (Claude Code subagents, Codex subagents). Scheduled jobs and wake-ups handle anything that depends on time or on someone else replying, so you stop watching for it (Claude Code scheduled tasks, Codex automations). The AI drafts every message, runs every site update and checks every report.

This is how my days look, with the setup I described at the top. When I want something done, I say so in the terminal and the AI picks the right connection.

Sign you're here: your first move on any task is to say it to the AI, and opening a separate app feels like a detour.

Hard gates stay human

What makes this safe is the gates. The AI drafts, and I say "send." It never sends, publishes, deploys or pays on its own. Anything that goes out to another person, goes live, costs money or deletes something waits for my yes. It's the arrangement you'd have with a sharp assistant who holds your keys, and you'd want it for the same reasons.

Guardrails matter more at this level than at any other, because more access means more ways for one mistake to travel:

  • Secrets never go into the chat.
  • Approvals come before anything goes out.
  • An agent's report is a claim. When a subagent says a job is done or a problem exists, check it yourself before you do anything you can't undo.
  • Rules live in a file the agent reads every session, so a correction made once holds from then on.

Which accounts an agent can reach and how much it may do on its own are your decisions. Base them on what you do and what you can afford to get wrong, rather than on a copy of someone else's setup.

The move: there's no Level 7. The work at the top becomes deciding what to ask for and judging what comes back, and every rule you write down makes tomorrow's sessions better.

Where to Go Next

If you're at Level 0 to 2, you don't need anything installed. Do the move for your level this week.

If you're at Level 3 and ready for the terminal, start with the setup guide for the tool you'll use: Claude Code setup or Codex setup. Both cover install and first run on Mac and Windows.

Past that, the official docs map every mechanism above: Extend Claude Code and the Codex docs index.

The fastest way to climb is to see what people one level above you are doing. The Daily Driver AI community is where members trade what's working at each level. Join us at jeangalea.com/daily-driver-ai, say which level you're on, and tell us what move you're making next.