Start with a repo you already understand
The fastest way to learn AI coding with Claude for free is to point it at code you know well enough to grade. Tutorials hide the failure modes. Your own repo does not. When the agent renames a function that three other files import, you will notice within seconds. On a toy project you will not notice at all, and you will walk away thinking the tool is better than it is.
Pick something with tests. Then ask for a change you could have made yourself in twenty minutes. Read every line of the diff. That single habit is worth more than a week of video courses.
The three weeks that actually teach you something
We run this progression with teams, and it works just as well solo:
- Week one: single-file edits only. Learn what the model sees and what it invents. Reject anything you cannot explain.
- Week two: multi-file changes with a test suite in the loop. Let the agent run the tests. Watch how it reacts to a red suite, because that reaction is the whole game.
- Week three: write your project instructions file. Most agent tools read a markdown file at the repo root for standing context. Put your build command, your test command, your directory conventions, and your hard rules in it.
A useful line to put in that file, kept blunt on purpose: test: npm test -- --run, and never add a dependency without asking first. If the agent cannot run your tests, it is guessing.
Free material that is worth your evening
Anthropic publishes its own documentation for Claude and Claude Code, and it is more current than any third-party course. Read the docs on context, on tool use, and on project configuration. Then read the changelog of whichever tool you use, because agent behaviour shifts release to release in ways that quietly break your habits.
After that, stop reading. The returns on more articles drop off fast. The returns on twenty more reviewed diffs do not.
What free learning will not give you
Being honest about the ceiling: solo practice teaches you to drive one agent on one codebase. It does not teach you what happens when six engineers point agents at the same service and the review queue triples overnight. It does not teach you where to draw the line on agent-written migrations, or how to keep a junior from shipping code they cannot defend in an incident.
Those are team problems, and they show up around month two of adoption, not week one. We built our training around them because that is where teams stall. If you are one person learning on a weekend, you do not need us yet.
Do this next
Open your own repo tonight. Give the agent one bounded task with a failing test attached. Read the diff line by line and write down every place it guessed wrong. Do that ten times over two weeks and you will know more about agent-assisted coding than most people who have finished a paid course.
When your team hits the coordination problems that solo practice cannot reach, that is the moment to bring someone in.