Whatever you are building, a turn has the same four beats.
1. You ask
One request, in plain English, about one thing.
Add a Clear button under the todo list that empties it.
Specific beats clever. You do not need magic words, a persona, or a template. You need to say what you want in a sentence you would be comfortable saying to a colleague.
The most common beginner mistake is asking for five things at once. Then when something is wrong you cannot tell which of the five caused it.
2. It plans
Before touching anything, a good turn starts with the agent saying what it intends to do: which files it will open, what it will change, what it thinks the tricky part is.
You can ask for this explicitly — tell me the plan first, do not change anything yet — and Claude Code also has a where exploring and proposing happen but edits are held back until you approve. Level 1 covers it properly.
Read the plan for one thing above all: did it understand the request? If the plan is aimed at the wrong file, you have just saved yourself twenty minutes for the price of reading four lines.
Quick check
The plan says it will edit a file you did not expect. What is the cheapest response?
3. It edits
Now it changes files. Depending on how you have set things up, it will either ask before each action or apply edits as it goes — Claude Code has modes for both, and level 3 is where you choose deliberately.
It may also run commands: your tests, a build, a script. It reads the output and reacts to it. That is not a gimmick; an agent that can run the test suite can catch its own mistakes before you ever see them.
4. You review
The agent shows you what changed as a : the lines it removed and the lines it added.
Read it. Every time. Three questions, in order:
- Is this what I asked for? Not “is it good code” — is it the thing.
- Did it change anything I did not ask about? Extra “improvements” are how surprises get in.
- Does it run? Start the thing, click the button, run the tests.
The honest failure mode
It will sometimes call a function that does not exist, import a library you do not have, or use an option that was never in that API. It will do this in fluent, confident, well-formatted code. The only reliable detector is running it — reading alone catches maybe half.
Then you go around again
Almost. The button clears the list but the counter still shows the old number.
That sentence is a whole turn. Short, specific, one problem. Three or four of these and the feature is done — and you understood every step, because you were in the loop for each one.
When the loop stops working
If you have gone around three times on the same problem and it is not converging, stop. Do not go round a fourth. Level 2 has a whole lesson on why starting over beats correcting five times, and it is one of the most useful habits in this course.
Do I have to review if the tests pass?
Is it faster to write one huge, detailed prompt?
Try it yourself
One full turn, all four beats. The last one is the easiest to skip and the most expensive to skip.