I’m using Cursor and I’ve noticed a behavior specific to Claude models that I don’t really understand: they often create one (or even several) Markdown files like SUMMARY.md, REPORT.md, etc., describing what they did.
For me this feels unnecessary and even counterproductive:
I’d rather get a short summary directly in the assistant message (a few bullet points), not in a separate file.
It wastes tokens on “documentation about the work” instead of the work itself.
With Haiku in particular, I sometimes see multiple Markdown files created for very small changes.
Other models I’ve tried in Cursor don’t seem to do this.
A recent example: Sonnet 4.5 Thinking created the following files after a change:
📚 Documentation Created
I've created 5 comprehensive documentation files:
NOTIFICATION_REFACTOR_SUMMARY.md - Quick overview and checklist
NOTIFICATION_SYSTEM_REFACTOR.md - Complete implementation details
NOTIFICATION_FLOW_DIAGRAM.md - Visual flowcharts with examples
NOTIFICATION_TEMPLATES.md - Template reference with formatting rules
NOTIFICATION_DEPLOYMENT_GUIDE.md - Step-by-step deployment instructions
I don’t need any of these files. They don’t provide value for my workflow and they’re just noise I have to delete manually.
Is this expected / by design in Cursor’s Claude integration? Is Cursor prompting Claude to “write a report” somewhere, or is it coming from Claude’s own defaults? And is there a setting or prompt tweak to disable this behavior and force a short inline summary instead?
This is not meant to happen, but used to be a problem with claude models. Do you have any skills/rules setup? If so, check if they mention any of this. What helps is creating a simple rule, saying the agent should not create summary md files. Hope this helps.
The model understands the context perfectly well without dozens of reports. Working with Claude, I had a situation where the model did nothing but write reports about what had supposedly been done, but in fact not a single line of code had been written.
I also had it happen many times with Haiku 4.5 specifically. Sometimes it just falls into the rabbit hole and keeps producing files like FINAL_REPORT_SUMMARY.md indefinitely - I had to stop it manually.
Hey, thanks for the request. This is a known issue with Claude models, and @liquefy suggested the right workaround.
Fix:
Create a rule in .cursor/rules/ that explicitly forbids creating summary files. For example, no-summaries:
---
description: "Prevent creating summary markdown files"
alwaysApply: true
---
Do not create markdown summary or documentation files (SUMMARY.md, REPORT.md, documentation.md, etc.).
Provide brief summaries directly in chat messages instead.
What else to check:
If you already have Rules or Skills, take a look there too. You might have instructions that tell it to generate docs. This behavior sometimes gets added by accident in custom rules.
About the issue:
The team is aware of it, but for now the best approach is the rule above.
In my experience, this propensity to create long MD files seems to be a quirk of Sonnet 4.5 (thinking and non-thinking). I didn’t experience it in earlier Sonnet versions, and in my cursory use of Opus 4.1 and my now daily use of Opus 4.5, I haven’t observed this behavior. I don’t have anything in my .cursorrules file prohibiting or promoting this behavior, which suggests that it’s a model problem. This was the primary thing that motivated my move away from Sonnet, although @deanrie ‘s workaround sounds like a good idea.
