Agents are generally too verbose in their chat responses and default to a report mode, when I need concise responses to my prompts. Reading through the unnecessary level of detail and human-unfriendly writing (long jargon-filled sentences, etc.) that agents dump on users is highly impractical.
However, enforcing writing rules to manage this is close to impossible.
Cursor injects built-in chat instructions you cannot edit. They steer answers toward report-style output and overly complicated answers. This means that any chat-specific writing rules I add sit in the same prompt and both apply at once, so the model gets contradictory instructions every turn.
I thought Hooks might be a solution, but unfortunately, chat replies are not part of that validation path. Unlike for other tasks, nothing checks the finished message before it appears and forces a retry.
The result is low, if any, compliance. The model may follow my rules at times, but mostly will ignore them.
Anyone having this issue? Any known workarounds? Is there perhaps a way to edit the built-in chat instructions I am missing?
You’re right that you can’t directly edit Cursor’s built-in chat instructions/system prompt, that’s not something we expose.
A few things that might help in practice:
Model choice matters. Different models weigh conflicting instructions differently. In my experience, my own rules reliably “win” over the default report-style steering, but it varies. Which models are you mostly using?
Phrasing matters even more. Firm, imperative rules tend to stick far better than soft suggestions. For example:
ALWAYS reply in short, concise answers. NEVER use jargon.
will outperform
Try to keep responses short and avoid jargon.
Front-loading ALWAYS/NEVER and being specific about what “concise” means (e.g. “lead with the answer, no preamble, no summaries unless asked”) usually boosts compliance a lot.
It’s contradictory with what you propose. Not sure if its because the model has changed since, or which is the most effective at this point.
But in both cases, it is very difficult to get cursor to comply to writing rules reliably.
Usually this leads to several re-prompts just to get formatting right. Which is a real hindrance, I have to say.