Cursor review - UI, Composer models, and where it still stumbles
I moved my daily coding into Cursor because I wanted the editor and the model in the same window, and I was tired of keeping a chat tab in the browser open next to a plugin that never quite matched my keybindings. A few months in, that bet mostly paid off: the UI was easy to pick up, the Composer models handled most of my daily tasks without me paying for the most expensive model every time, and I still lost work when the app froze and had a harder time with skills and rules than I did in Claude Code.
The UI was easy to pick up
Cursor is a VS Code fork, so the layout was already muscle memory with the file tree on the left, tabs up top, and the terminal where I expected it, and I did not spend a weekend remapping keys or hunting for where they hid the debugger.
The AI pieces sit in places that made sense to me after one afternoon, with chat in a side panel, inline edits showing up in the file with accept and reject buttons, and Composer mode opening when I want multi-file work instead of a single highlighted block. I still reach for the command palette for odd jobs, same as before.
What helped me most was how little friction there was between typing and asking, because I can highlight a function, hit the shortcut, and send the question in with the selection attached instead of copy-pasting paths into a browser tab. I live in the editor most of the day, so that small loop saves real time compared to bouncing between tools.
| How I work | Where it lives in Cursor | When I reach for it |
|---|---|---|
| Quick question on one function or file | Chat side panel with selection attached | Most of the day when I am reading code |
| Small edit in place | Inline accept/reject in the editor | Renames, error handling, local fixes |
| Multi-file pass | Composer | Refactors, tests, touching several paths at once |
| Odd command I do not remember | Command palette | Same as VS Code, unchanged habit |
Composer models are good for the price on everyday work
I use Composer models for most of my day on refactors, test stubs, and walking through an old module I did not write, and they score lower than the top names on public leaderboards even though they have been good enough for what I pay in the subscription that I stopped opening a separate premium chat for routine tasks. I used Composer 2 for a while before 2.5 landed, and 2.5 felt like a real step up on the same kinds of work: multi-file edits hung together better and I reached for a heavier model less often on tasks that used to borderline need it.
Short answers come back fast, while bigger multi-file passes take longer in a way that makes sense when the tool is reading a lot of files at once. For boring chores like renaming a prop across components or drafting unit tests from an existing function, quality has been solid, and I still double-check anything that touches auth, payments, or schema migrations because Cursor did not change that habit.
On hard greenfield design or subtle concurrency bugs I switch to a heavier model or think longer myself, but for normal daily grunt work Composer has been worth the money.
| Kind of work | Model I reach for | How it has gone for me |
|---|---|---|
| Refactors, renames, test stubs | Composer 2.5 | Step up from Composer 2 on the same chores |
| Auth, payments, schema migrations | Composer 2.5, then my review | I still read every line before merge |
| Greenfield design, concurrency bugs | Heavier model | Composer is where I notice the gap |
Freezes and crashes still bug me
The main problem is stability, because Cursor has frozen on me mid-session with the UI locked up, the spinner spinning, and sometimes the whole window going white until I force-quit. I have had weeks with no trouble and then two bad days in a row where I lost a Composer thread because the app died before I copied the last reply.
I save often and I commit before big agent runs, which feels like extra homework for a tool I pay for every month. Some weeks I see no crashes, and other weeks I get two in a row and hesitate before kicking off a long edit across twenty files, which is why I keep a second editor installed.
Skills and rules are weaker than in Claude Code
I wanted project-level guardrails around naming conventions and folders the agent should never touch, and Cursor has rules files and skills support where I got basic behavior working with lines like "use our logger" and "do not touch generated protos," but the docs and examples still lag behind what I found in Claude Code.
| What I set up | Cursor | Claude Code |
|---|---|---|
| Project rules | .cursor/rules files; basics worked |
Clearer examples when I was learning |
| Reusable skills / workflows | Improving; agent sometimes ignored a rule until I pasted it again | Easier chaining and sharing across repos |
| Checklists the agent re-reads | Possible, more hand-holding | Easier to bake in from the examples |
| Skills marketplace online | Smaller selection; more DIY | Much more developed; easier to browse and install |
| Docs when you are new to it | Still catching up | What I leaned on first |
In Claude Code I found clearer examples for chaining skills and sharing them across repos, and the skills marketplace online is much more developed than what I have found for Cursor so far, with more ready-made skills to browse and install instead of writing everything from scratch. and checklists were easier to set up so the agent actually re-read them. Cursor improved while I was using it, but I still had times where the agent ignored a rule until I pasted the same line into the prompt again, and I can work around that with more hand-holding than I wanted after writing the files once.
If your team already standardized on Claude's skills layout, moving to Cursor means rewriting that setup, while if you mostly work alone and only need a short rules file Cursor is probably fine. I care about reuse, and I want Cursor's skills and rules to catch up.
Would I keep paying?
Yes, for now, because the UI is easy to use and Composer models cover most of what I do without me upgrading every task to the priciest model, even though I keep another editor installed for the days Cursor locks up and I still read what the agent writes before I merge.
If Cursor stops freezing as often and skills and rules work as reliably as they do in Claude Code, I would tell friends to use it without caveats, and for now it is still the editor I open first every morning.