Composer 2.5 has lost the plot

Support Help
,
1

First I do not programme for a living so forgive what might be a daft question. I have been using Composer 2.5. At the start of the project things went well now after 10 days Composer 2.5 is making lots of mistakes and not conforming to rules I set. Is this typical and is there anything I can do about it? Or can anyone suggest a solution?

Thanks in advance.

J

5

Hey, that’s not a dumb question at all, and what you’re describing is totally expected.

It sounds like the whole project ran in the same chat for 10 days. When one chat gets really large, it hits the context limit and starts auto-compressing the history into a summary. After that, the model is working from the compressed summary, not the original messages. So over time it loses early details and starts following rules less consistently. It’s not a model bug and it’s not something wrong with your setup, it’s just how a very long session behaves.

What actually helps:

  • Start a new chat for each task instead of keeping everything in one giant session. Sounds like you already started doing this a few days ago, and that’s the right fix, not switching models.
  • Move important, always-on instructions into Project Rules at .cursor/rules and keep them short. They get injected into every request, but following them is probabilistic and gets weaker in an overfilled chat. Rules docs: Rules | Cursor Docs
  • At the start of each new chat, quickly restate the critical constraints and clearly point to the key files or specs you’re working with.

Try this for a couple of tasks and tell me if it feels better.

6

Hi Dean
Many thanks for the feedback. I got excited at the start if this project now I have calmed down I read the docs and any cheat sheets I could find so life has got a little less fraught.
I did add some more money as I ran out in the first month. Is it me or does money added seem to burn faster than the subscription?

Once again thanks for taking the rime to reply

7

Glad it feels a bit easier after reading the docs.

About the money, added funds on demand are charged at the same API rates as the usage included in your subscription, so the rate itself isn’t higher. If it feels like they burn faster, it’s usually because of two things:

  • Which model you’re using. Composer 2.5 and Grok 4.5 come from a separate included Cursor Models pool and are cheap. Third party models like Claude, GPT, and Gemini are billed at the provider’s prices and will use funds much faster. If you switched to something like claude-sonnet-5-thinking, that can speed up spending a lot.
  • Context size. The longer the chat and the more context in the request, the more tokens each request uses, so the cost goes up. That’s also a good reason to start a new chat per task, it’s usually better quality and cheaper.

You can see an exact breakdown per request on the Spending dashboard: https://cursor.com/dashboard/spending. It shows the model, the cost, and which pool it came from.

If you see something there that looks like a billing mistake for your account, send a note to [email protected] and they’ll check your specific case.

8

Hi Dean

Yet again great advice thanks. I am also using plan and auto now on a regular basis and opening new chats.

Thanks for all your advice…