The mistakes people make in week one

and then I asked it to "just handle auth" in one prompt, no error pasted, no file named, and I was shocked when the answer was a lecture plus a rewrite I couldnt map back to my repo. That was day two. Day three I opened a different tool to fix the rewrite. Classic.

Week one is where people decide these things are magic or useless. Both takes are usually the same five mistakes. I have done all of them. I will do two of them again this month if I am in a hurry.

The first one is the giant prompt.

You know the shape. Fix the login, clean the spreadsheet, and also finish the landing page, and make it production ready, and explain every choice. It feels efficient. You are batching. What you get is a long answer that is slightly about all three and useful for none. The login part invents a middleware you do not have. The spreadsheet part invents column names. The landing page sounds like a template you have already closed a tab on.

I did a version of this on a tuesday with a 401 and a CSV in the same message because I did not want to "waste" a chat. Waste compared to what. I spent an hour untangling a reply that mixed my route names with a generic Next example. Split the asks. One sentence each, like I said in you dont need another tool. The login is one chat. The spreadsheet is another. If you are in Cursor, start a new chat when the task changes. Their agent docs say that. I ignored it and then complained the model was confused.

Smaller prompts feel slower. They are not. You get something you can accept or reject. That is the whole game.

Second, not pasting the error.

I still catch myself describing a bug like I am at a standup. "Login is broken." "It just spins." "The dates are weird." The model cannot see your network tab. It cannot see the cell format. It will guess, and the guess will be the most common bug in its training, which may not be yours.

Paste the thing. Status 401, response body, the two log lines, the screenshot if that is all you have. In Cursor, open the folder and point at the file, the Agent can search but it still does better when you name the route. In Claude, upload the export or the draft, do not summarize the spreadsheet if you have the spreadsheet. In Grok, paste the policy email or the booking code plus what the site showed you.

The half-written landing page is the same mistake in nicer clothes. People say "make this better" and attach nothing. Better than which draft. Paste the draft. Even the ugly one. Especially the ugly one. I once asked Claude to finish a pricing section from memory of my own copy and it wrote a cleaner paragraph that dropped the only sentence I cared about. My fault. The sentence was sitting in another tab.

Third, switching tools mid-problem.

This one looks like research. Cursor gives you a diff you do not like, so you paste a fragment into Claude. Claude's rewrite sounds smarter, so you take that to Grok to "see what it thinks," and now you have three answers and one broken login. I have a notes file that is just competing patches. None of them applied cleanly because each model was missing a file the others had seen.

Pick with which tool when and stay there until the sentence is done or clearly in the wrong place. Wrong place is fine. Move the same sentence, the same error, the same draft. Do not move a vibe.

A fair switch: you opened Grok for a current pricing question and it turned into a 400 line code change you need in the repo. That is Cursor's job. Take the requirement, not the generated file tree, and start a Cursor chat. Other fair switch: Cursor keeps missing the tone of an email you have to send. Take the facts to Claude and write the email there. Unfair switch: you are bored and a friend likes a different logo.

Fourth, trusting the first answer.

They all sound sure. That is the product. Cursor will apply edits as it works and show you a diff, which only helps if you read the diff. I have accepted a login "fix" that also loosened a check I needed. Technically I could sign in. So could a slightly wrong password, for a minute, until I noticed. Claude will hand you a polished email with a date it inferred. Grok will cite a page. Open the page.

A thing I do now that I should have done week one: run it, or read it out loud, or send it to myself. For code, actually hit the login with the test user. For the spreadsheet, sort the date column and see if January lands where January should. For the landing page, read the pricing section like you have no idea what the product is. First answers fail that test a lot. Second answers fail less, if you say what was wrong.

I also ask what it changed. In Cursor that can be literal, walk me through the diff, what did you touch besides the handler. In Claude, quote the lines you added that were not in my draft. In Grok, list the claims that depend on a link. If it cannot, I get suspicious. Not mystical, just tired of confident empty calories.

Almost right is the dangerous one. It passes a skim. It fails at 11pm when you trusted the skim.

Fifth, never saving a working prompt.

You will get one that works. "Here is the 401 body, only touch login.ts, do not invent middleware, write a test that hits the test user." Then you close the chat and next week you type fix auth again. I did this for a month. I could not remember the prompt that made the landing page pricing section acceptable. I remembered the feeling of it.

Put the good prompt in a note. Not a second brain system. A note. Date it. Paste a before and after if you are proud of it. Cursor users eventually grow rules and I will let the getting started series talk about that when the docs still say so. Week one you need a note. Claude Projects can hold the messy source files so you stop re-uploading, that is a form of saving. Grok threads can be continued, so do not start a fresh "what changed" chat if last Tuesday's thread already has the airline name and the booking.

A working prompt is more valuable than a new account. I say that as someone who still has extra accounts.

A couple smaller ones, since week one is generous with ways to waste a night.

Accepting a huge rewrite because it "worked." If the login needed a header and you got a new auth folder, revert and ask smaller. Cursor can restore a checkpoint from an earlier message. I use that when I have been lazy.

Doing the comparison homework instead of the problem. I already wrote which tool when. If you are on day four and still deciding, you are collecting. Open one. Cursor if its files. Claude if its a pile of words. Grok if you need the current version of a public fact.

Week one does not need a system. It needs one sentence, the error or the draft pasted in, and you staying in the room long enough to check the answer. When that works once, save the prompt. When it fails, say what failed, in the same chat, before you go shopping.

I still write giant prompts when I am annoyed. I still forget the error paste. The difference now is I notice around minute five instead of minute sixty, and I am slightly more embarrassed, which helps.