How To Use Cursor With Claude For Beginners

Source: YouTube auto-generated captions, lightly cleaned (repeated caption lines removed, misheard brand names corrected).

0:00 — Wasted spend, no negatives

This morning, one live account spent $875 in the last 30 days. $182 went to search terms with zero conversions. That is 88 distinct terms that should already be negatives. Average CPA sits at $73. Buddy is the agent that is supposed to catch this. Buddy itself was built in Cursor with Claude. Cursor is a code

0:25 — Chat will not ship Buddy

editor. Claude is the model inside it. Beginners treat that pair like a chat window and paste snippets by hand. That is slow, and it is how files drift. Buddy was not built that way. The same four moves that shipped Buddy are the ones you should use on your first

0:44 — Why this order matters

project. The order is agent mode, then {dot} cursor rules, then a browser audit, then ship. Agent mode first, so Claude can edit the repo instead of talking about it. Rules next, so it does not invent a second style. Order before ship, so tracking is not a guess. Ship last because wasted spend only shows up

1:06 — Step 1 — Open Agent mode

live. Step one is agent mode, not the chat panel. Open Cursor, switch the composer to agent, and point it at the Buddy repo. Agent mode can read files, edit them, and run the terminal. You do that first because a chat answer you paste by hand will miss half the files Buddy actually needs. Give agent one job, not five. For Buddy, the first job was load the account, list search terms, and flag zero conversion queries. Keep the prompt short, so Claude stays inside the repo.

If you stack unrelated tasks, agent mode starts rewriting working files, and you spend the hour putting Buddy back together. Step two is a {dot} cursor

1:48 — Step 2 — Add .cursorrules

rules file at the root of the project. This is a plain text list of how Buddy is allowed to work. You write it before Claude generates more code because rules are cheaper than undoing a wrong folder layout. Buddy's rules name the stack, the folders, and what never gets touched. Keep the rules short. Stack, folder names, test command, and do not touch paths. For Buddy, that meant ads data stays read-only until a human confirms a negative.

Long rules get ignored. Claude follows the first screen of constraints, then improvises. If a rule does not change a decision, cut it so agent mode actually obeys the rest. Step three is a browser audit of

2:32 — Step 3 — Run browser audit

the running app, not a code review. Start Buddy locally, then have Cursor open the page and click the same path a media buyer would. You audit in the browser because ad software fails in the UI. A green terminal does not prove the wasted $182 would have been caught. Check three things on that page: the spend total, the zero conversion list, and the proposed negatives. If any number does not match the account, stop.

Buddy showing $875 spend and $182 waste is only useful if those figures are the live ones. Wrong totals in the UI mean you ship a liar. Step four is ship, not another

3:16 — Step 4 — Ship the build

local tweak. Commit the files, deploy Buddy, and point it at the live account. You ship because a local demo never sees the next day's search terms. The $182 waste and the 88 negative candidates only matter if Buddy is running against the account tomorrow morning. Shipping is not a big bang release. Push a thin slice. Read search terms, list zero conversion queries, draft negatives. Leave apply negatives behind a confirm. That is why Buddy can sit next to a 73 CPA account without writing keywords it should not. Ship the read path first so you can trust the right path later.

Once that slice is live, the morning

4:00 — Once the slice is live

numbers become a checklist. $875 spend, $182 on terms with zero conversions, 88 negative candidates. Buddy's job is to surface those 88 before a human has to export a search terms report. That is the only reason the cursor work existed. If Buddy misses them, the build failed. The beginner trap is skipping rules in the browser,

4:26 — The beginner trap

then asking agent mode to ship. Claude will write a lot. None of it will match the account. You will spend the afternoon comparing 73 CPA in Google Ads to a different number on a local page. Do the four steps in order so that comparison never happens. Agent mode so Claude can edit the repo. Dot cursor rules so it does not invent a

4:48 — Recap

second Buddy. Browser audit so $875 and $182 are the real figures. Ship a read-only slice so 88 negative candidates show up tomorrow without writing keywords. That order is how Buddy got built. Skip a step and you rebuild it. If you are new to cursor with Claude, copy the Buddy sequence and keep the scope small. One repo, one job, four steps. The full write-up of this path is on the blog. The blog is on ahmeego.com.

Buddy by Ahmeego, a company of It All Started With an Idea.