How I'd start over with AI coding if I were starting today
5 min read
The advice for someone starting AI coding in 2026 is different from the advice we would have given in 2023, and we are past the early-adopter phase enough that there is now a clear right way to start. Here is the retrospective version: what we would do if we were doing it again.
Pick one tool and commit
The biggest mistake new users make is tool-shopping. They install Claude Code, then read about Cursor and install that, then hear about Windsurf and try it, and a month later they have shipped nothing because they have been comparing options instead of building.
The tools are close enough in capability that any of the top four works. Pick the one that runs on the AI you already pay for, or the one your friend uses, or the one with the documentation that reads best to you. Then stop researching. Build something with it.
A weekend of focused use will teach you more than a month of reading about which one is best.
Build three projects in your first weekend
Not one. Three. Because the first one is where you learn the tool, the second one is where you start to relax, and the third one is where you realize the skill transfers.
A reasonable trio for a non-developer: a personal site, a small script that does a chore you do regularly, and a tiny web tool with one input and one output. Each fits in an afternoon. Each ends with something you can show, or use, or both.
If you build only one project, you will attribute the experience to the project. If you build three, you will see the pattern, which is the actual lesson.
Learn the loop, not the syntax
Every AI coding tool shares a five-step workflow. Ask. Review. Approve or correct. Checkpoint. Repeat. That loop is the skill. Everything you might think of as "technical knowledge" sits inside it, and the AI handles most of what falls inside.
When you sit down with a project, your job is to brief the AI well, glance at each diff, give plain-language feedback when something is wrong, and commit checkpoints when something is right. That is it. That is the entire craft, and you can be fluent in it inside a week.
The temptation, especially for people coming from a "I should learn to code" mindset, is to try to learn the language the AI is writing in. Don't. Read the diffs for shape, not syntax. Let the AI handle the parts you don't read yet. The depth comes later, by exposure, and only if you want it.
Resist the framework rabbit hole
Within a week of starting, you will hear about frameworks. React. Next.js. Tailwind. Svelte. Vite. People will tell you which one is best. Online posts will declare one of them dead.
For the kind of small project you will be building in your first month, all of these are fine, and the AI will choose a sensible one for you. The framework matters far less than the marketing suggests. Let the AI pick, build with what it picks, and worry about the choice later, if at all.
Six months from now, when you are building something larger, you will have the context to make the choice yourself. Spending your first weekend on framework comparisons is a known way to ship nothing.
Treat the documentation as a friend, not a destination
Every AI coding tool has documentation. It is worth knowing exists. Skim the quickstart. Bookmark the commands page. Then close it.
The mistake is treating documentation like a textbook to read cover to cover before you are allowed to start. The right model is: build, hit something you don't understand, search the docs for that specific thing, return. The reading-as-you-go habit is what makes the documentation useful. The reading-before-starting habit is what makes you bounce off after two days.
Build something useful, not impressive
The single best heuristic for the first month is to build things you will actually use. The quote calculator for your business. The spreadsheet summarizer for the chore you do every month. The dashboard that shows you the three numbers you actually check.
The pressure to build something impressive, something that would make a good social media post, is the enemy of habit formation. Useful things keep you coming back. Impressive things sit in a folder.
Six months from now, when "useful for you" turns into "useful for someone else, who might pay for it," you can start building impressive things. By then you will have the reps.
The whole arc in one sentence
Pick one tool, build three useful small projects in your first weekend, learn the loop cold, stay away from framework arguments, and treat the documentation like a friend rather than a destination. That is it. That is the whole on-ramp.
Frequently asked questions
How long until I'm good at AI coding?
Productive in a weekend. Fluent in a month of evenings. Capable of shipping something a stranger would pay for, six months to a year. The on-ramp is much shorter than learning to code traditionally; the depth still takes time.
What's the single most important thing to learn first?
How to write a tight prompt. Everything downstream gets easier when the brief is specific. The second most important: the habit of committing every time the project is in a good state.
Should I learn to code traditionally before using AI?
No, unless you want to. The AI handles the typing. What you need is enough literacy to read diffs and spot when something is off, which you'll absorb by exposure during the first few projects.