Pick one real problem

and then I typed "build me a small app that handles login and a dashboard" because I wanted to learn Cursor, which is how you spend three hours reading diffs you did not ask for.

That was night two. Night one I at least had a 401. Night two I had a syllabus.

This is step 2. Step 1 was getting the editor open. If you still do not have a sentence, You dont need another tool is the earlier lecture. This one is about picking work that can actually finish.

Not "learn Cursor"

I see people treat the first week like a course. They ask it to explain Agent vs Ask vs Plan. They ask it to write a todo app. They ask it to generate a portfolio site so they have "something to show." Then they quit, because the output is a pile of files they do not own, and the next real bug still sits in the old repo.

Cursor is an editor with an agent in the side panel. You already know how to use an editor. The new part is asking it to change a thing you can check.

So pick a problem that existed before you downloaded anything.

A login 401 with a known test user.

A script that dumps a CSV in the wrong column order.

A hero section that will not center on mobile, the one you have already cursed at in Chrome.

A test that fails on Tuesdays because of a date helper.

Those are problems. "Get good at Cursor" is not a problem. It has no done.

Small asks, on purpose

Their own quickstart says start with a small task after you pick a folder. Their example first change is the kind of thing you can see in a browser. Mine, when I am not being an idiot, looks like this.

The contact form button does nothing when email is empty. Here is the component. Do not add a new form library.

Or: this script writes name, email and I need email, name, created_at. Keep the same filename.

Or: the heading on /pricing sits left on a 390px window. I already tried text-align: center on the wrong div.

Notice I am not asking it to redesign the site. I am not asking it to "make auth solid." I did that. It rewrote half the auth file. I said that already. I will keep saying it because I still catch my hands typing the big version when I am tired.

A small ask has a place you will look when it is done. The button shows an error. The CSV has three headers. The heading is centered. If you cannot name the place, the ask is still a vibe.

I write the ask in the same note I used for the one-sentence problem, then I paste it into Agent with Cmd+I. Sometimes I @ the file. Sometimes I just describe the route. Both work better than "can you look at this project and improve it."

Why "build me an app" is how people quit

Week one energy is dangerous. You have a new editor. The panel is empty. It feels wasteful to ask about one CSS file.

So you ask for a whole product. Agent is allowed to do that. It will search, edit multiple files, run terminal commands, keep going. Official help is pretty proud of that. I am not arguing with the feature. I am arguing with using it as your introduction.

A generated app is a stranger's repo that happens to live on your disk. You did not make the folder decisions. You cannot explain the auth. When something breaks, and it will, you have no map. Then you decide Cursor is sloppy, or you decide you are bad at it, and you go back to fighting VS Code, or you open Claude and paste one file like you should have done in the first place.

I watched a friend do the full-stack week one. Friday he had a dashboard. Sunday he could not change the nav color without the agent rewriting the layout. He uninstalled. The nav color was the real problem. He should have started there.

The app-sized prompt also hides the only skill that matters, which is reading the diff. A 20 line change you can reject. A 400 line scaffold you just accept because reviewing it feels like a second job.

If you truly have no existing repo, fine, make a single page. One HTML file. A script that renames a batch of photos. Do not start with users and sessions and Stripe. That is how the first tuesday ends with a new lib/auth.ts you have never opened.

A few asks I have actually used

The reset password link in the email pointed at localhost. I pasted the email template and the env file name, not the secrets, and said production should use APP_URL. That was a real hour. It tried to add a whole config module. I rejected that part and kept the one line.

The landing hero would not sit in the middle of the viewport because a parent had align-items: start. I told it the page, the breakpoint, and that I did not want a new layout component. It still offered a wrapper. I said no, just the parent. That is a legal conversation. You can talk back.

A node script that merged two CSVs kept dropping the third column when the name had a comma. I pasted the error and two sample rows. I did not say "rewrite this in pandas." I said keep it in node, fix the parse.

None of those taught me Cursor. They taught me whether I could check the result. That is the week one test.

If your problem is a long email or a pile of notes, you are in the wrong tool tonight. That is Claude. If you need a current "is this still how people do it" answer, that is Grok. Cursor when the work is already a folder, which is the rule I keep repeating in Which tool when.

How I know I picked right

I can say done in one sentence. "The test user can log in." "The CSV has the three headers finance asked for." "The heading is centered at 390px."

I can point at one or two files I expect it to touch. If I cannot name a file, I still have a problem, I just have not looked yet. That is fine. Open the folder, search yourself for thirty seconds, then ask. The next post is show it the files because a good problem still fails if you hide the repo.

I am willing to reject the extra work. If I asked for a null check and it refactors the helper, I do not keep the refactor as a bonus. Bonus refactors are how you lose the afternoon.

If the ask is still "help me get comfortable," shrink it. Take the ugliest thing on your disk and make one part less ugly. Then stop for the night. When it goes wrong is waiting when you do not stop. Make it stick is the habit after you have had one win.

Theres no prize for using every mode in week one. There is a prize for a smaller bug than you had this morning.