Composer 2 performance seems quite degraded.
Has anyone else had the same experience?
Hi @gabriel-filincowsky Thanks for your post! Do you mind providing a bit more detail? Is it slow? Not answering the question as well as you thought? Is it rushing to answer without completing the task?
Feel free to provide a request id for a request in which you notice the poor performance (with privacy mode ideally disabled please)
To disable Privacy Mode:
- Open Cursor Settings with
Cmd+Shift+Jon macOS orCtrl+Shift+Jon Windows/Linux. - Go to
General. - Turn
Privacy Modeoff / switch toShare Data.
To get the Request ID:
- Open the relevant conversation in the Chat sidebar.
- Click the
...menu. - Select
Copy Request ID.
Hi Kevin, thank you for your interest and for the question. It is a bit difficult to clearly describe everything I have been observing with Composer 2, but I will try to summarize it without going into too many examples so it remains readable.
In my case, Composer 2 has never been the most reliable model for deep reasoning on complex topics outside of coding. I should clarify that my use of Cursor is mostly not coding-related. Roughly 80% of my work is mechanical and chemical engineering, and only about 15-20% is coding. Because of that, I often use it for structured tasks (well-defined and not content-dependent), and in general, it has worked reasonably well for those workflows.
What I have noticed recently, over roughly the past week or so, is a decline in instruction following. Instead of strictly following the given steps, the model tends to reinterpret the instructions and act based on its own inferred intent. I usually prefer the interpretation of intent by the model. However, in this case, it is less reliable and less aligned with my intention or requirements.
This becomes especially problematic in workflows where the steps are explicit and mechanical. For example, I have used it to convert PDFs to Markdown and to clean up already-converted Markdown files. These workflows used to work fairly well, but more recently Composer 2 has started deviating from the instructions, skipping steps, or changing the output structure. In one case, when I asked for a simple optimization of an already-converted Markdown file, it attempted to introduce external commands and temporary scripts, which ended up corrupting the file structure entirely and shuffling the content in an unusable way.
I then tried again in a fresh session using a backup version of the file and explicitly instructed it not to introduce scripts or modify the workflow beyond the defined steps. Even then, it did not consistently follow those constraints.
Another thing I have noticed is a change in formatting behavior, particularly an overuse of bold text. In some outputs, bold formatting is applied excessively, most of the time on individual words rather than full phrases, and often in contexts where it is not necessary at all. This results in a much heavier emphasis style than intended, sometimes affecting a large portion of the text.
Overall, what I am seeing is a reduction in instruction adherence and consistency. For structured or repetitive tasks where I rely on clear, predictable behavior, the model now feels less reliable. In my case, that leads to outputs that are less aligned with the intended workflow.
I would love to have an efficient model like Composer 2 that can also handle more complex tasks in my specific domain. I fully understand and respect the decision by the Cursor team to focus it on coding, which makes complete sense. However, even in coding-related documentation, it would be useful if the model could produce clear, well-structured writing, because at the moment the explanations it generates are often difficult to follow, with weak logical flow and some inconsistencies. This becomes even more noticeable when I try to use it outside of coding, where the performance tends to degrade further.
I asked composer 2 “why not following the rules/steps even I wrote down ?” guess what the answser 
composer suggest that using rough/command words like “must” or something like that.
Then the interesting part is that I ask composer: What’s the point of writing it down if you’re not going to read it?
Conclusion: No matter how many words you use to describe it, or how precisely you describe it, the composer either doesn’t read it or ignores it, and then freely interprets it according to its own understanding.
I’m with you on this, I have a set of rules created as the doc says, using componser to create said rules with the doc, and it’s impossible to make him follow any of the rules. Tried creating new sessions, making them short, quoting the exact rules I want to follow. Ignoring also my “musts” like always ask before editing code.. it’s annoying and I’m tempted to finish the pro payment
It’s always a bit subjective, but I also get the impression that Composer 2 is currently doing everything except what it’s supposed to.
Yes looks like Composer 2/Kimi 2.5 is being trained like that, it just weighing internally about performing the task in the user’s instructed way or its trained way. Most of the time I need to make my skill’s instruction very long and detailed, step-by-step, like a massive amount of tokens just for the skill’s instruction so Composer 2 can follow it end-to-end, and it’s still randomly processes the task on its own. Very frustrated if you keep on expecting it to follow the instructions.
I’ve observed and found that Claude and OpenAPI models only tend to not follow the instruction on a long session when it runs for a while.
Thank you for the thoughtful feedback. I’ve shared this with the team. We’re always working to improve Composer, and this kind of feedback helps us tremendously. Stay tuned for updates.