How I'd Learn AI From Scratch in 2026

Chat, delegate, build. Start with a job you actually need done.

AI Confidence Accelerator: Chat, Delegate, Build. 30 lessons, lifetime access, with a preview of the course.

Want to work learn AI with me?

Join the AI Confidence Accelerator waitlist for first access and 25% off at launch. Thirty days, about twenty minutes a day, covering Chat, Delegate, Build and Teach.

If I was starting to learn AI today, I'd work through three stages: chat, delegate, build.

I spent Friday's livestream walking through the lot. Step by step. What to actually do, which tools to start with, and enough of the concepts to understand what's happening. You can watch the full 55-minute session here, or pick the shorter lesson for the stage you're at below.

But please do something with it! Watching another twenty tutorials can become a very comfortable way of avoiding the thing you're trying to learn. You cannot learn to drive by reading a book. You have to get your hands dirty and DO the thing

1. Chat

We all start here - with the basic chatbot. Working out what AI can do and how it fits into our life.

Pick ChatGPT or Claude and get comfortable using it. Which you choose is a terribly uninteresting question. They are both good.

Give it a messy idea. Ask it to challenge you. Show it the spreadsheet, dictate what you're thinking, take a photo of whatever has confused you. Learn the limits of what AI can and cannot do so that you can better imagine uses. A lot of limitations with usage come from not realising what it can do.

For ongoing work, create a Project with the relevant files and instructions. I used a Claude Project to help work through my move to Cyprus, including the residency paperwork. Keeping the background together saved explaining it every time.

When the results are vague, check the brief. RISEN helps: Role, Instructions, Steps, End goal, Narrowing. Who's it helping? What should it do? What does a good result look like? Give it an example. You're practising communication.

RISEN: Role, Instructions, Steps, End Goal, Narrowing.

2. Delegate

This is where we stop going back and forth with AI and instead delegate jobs to AI. We handoff and (ideally) never have to worry about that particular job again.

Find a job you repeat each week and can judge yourself. A briefing, a client update, collecting invoices. If you can't think of one, ask AI to interview you about your working week.

An agent can use connected tools to carry that job out. Start with the desktop app, such as ChatGPT Work or Claude Cowork, and check what access you're granting when you connect your files or email.

Use Select, Scope, Supervise, Sign off. For a weekly briefing, specify the tools you care about, the date range, original sources and where the finished document should go. Check the first run. If it sends you Claude news when you use ChatGPT, tell it and update the instructions.

Delegate with 4S: Select, Scope, Supervise, Sign off.

Once it works, save the process as a Skill and schedule it. Keep approval on client messages or anything involving money! This newsletter is an example: AI prepares the draft from my livestream, then I edit it manually.

That saves me 10+ hours a week easily. You need to find similar tasks for yourself.

3. Build

This is where we shift from handing off jobs to actually having AI build tools for us and then our customers.

You’ll need to shift from ChatGPT and Claude to Codex and Claude Code at this stage. First step is purely to open a project folder in Codex or Claude Code and describe a small tool you'd use yourself. Use natural language - just tell it what you want to build and whoosh off it goes.

Get the first version working. Then try to break it. Click things in the wrong order. Give it to somebody else; they'll find a problem you missed almost immediately. Show the AI what happened and what you expected, then test the fix.

When you're ready to share an app, learn how to protect its data and control access. My stack uses GitHub for code history, Vercel for hosting and Supabase when I need a database. You can learn each piece when your project needs it.

Kyle's tech stack: Codex or Claude Code, GitHub, Vercel and Supabase.

Learn it with me

If you've been using AI but still feel like you're winging it (that’s not unusual!), I've been building something for you: the AI Confidence Accelerator.

Stop winging it. AI Confidence Accelerator: Chat, Delegate, Build and Teach.

It's 30 days, about ~30 minutes a day. We work through Chat, Delegate, Build and Teach, with one lesson each day. You'll practise giving AI proper briefs, handing off real work and building your own tools. Then you'll explain what you've done to someone else, in your own words, and answer their questions.

Actual AI Confidence Accelerator interface showing Day 3, Feed it your world, and the course navigation.

Inside Day 3: working with your own documents.

Actual course exercise: record or type an explanation, save it and receive feedback.

Record your explanation and get feedback.

I want you to know what the hell you're talking about when your client, boss or colleague asks! Join the free waitlist and I'll email you when the programme opens, with a code for 25% off at launch.

To the Task,

Kyle