This .md file creation is absolute garbage output. I know what changes or code creation i ask for I dont need verbose self congratulatory output like this - who ever thought this to be a good idea was wrong
Can this be deactivated or does it need to be through a set of rules (and potentially another .md file) ?
I don’t really use Claude much, though I remember he likes to create extra temporary files. Maybe when I work with him, the chance of separate documents being made for each task is lower, since I’ve been channeling his reports straight into the chat.
1 Like
sometimes， the claude would create .md and then delete
1 Like
Where does the bug appear (feature/product)?
Cursor IDE
Describe the Bug
I have a rule to NEVER automatically create .md files. Just got 5 .md files created on just 1 request.
Steps to Reproduce
I have no idea. Try to add a rule to never create .md files and ask for a new feature?
Expected Behavior
no .md files generated.
Screenshots / Screen Recordings
Operating System
Windows 10/11
Current Cursor Version (Menu → About Cursor → Copy)
Version: 1.7.44 (user setup)
VSCode Version: 1.99.3
Commit: 9d178a4■■■89981b62546448bb32920a8219a5d0
Date: 2025-10-10T15:43:37.500Z
Electron: 34.5.8
Chromium: 132.0.6834.210
Node.js: 20.19.1
V8: 13.2.152.41-electron.0
OS: Windows_NT x64 10.0.26100
For AI issues: which model did you use?
Sonnet 4.5
For AI issues: add Request ID with privacy disabled
23c50b51-b9a2-46ee-a02e-1b71448e1ff5
Does this stop you from using Cursor
No - Cursor works, but with this issue
When I told him not to create md files, he started doing the same thing, but in txt files. LOL.
3 Likes
so smart
I think Sonnet 4.5 might be the culprit! I’ve noticed the same issue with Copilot, Rovo Dev, and a few other tools. It can be pretty frustrating! It makes me wonder if there’s a conspiracy (
) at play, possibly to encourage us to spend more
1 Like
Same issue here… im on Auto and .md files are generated on each request.
Still happening even with project rules to not write md files
Same issue, the most frustrating thing, on top of paying 4x the price with bigger context because 50% of the context is basically lost to this garbage. So on top to the price you are paying vs. the price advertised is very far apart, you are getting sub standard claude. No rule stops this.
Only thing a rule does is, it makes it delete the md file after creating it. Context, tokens and time still wasted.
It loves da md files, I would argue some micro level md files are useful to track apps evolution, where does this data normally reside in JIRA? I am one man shop so don’t operate like a big org. It would be good if they are added to memory, persistent memory needs to be a thing with AI models. E.g. AI ‘wakes up’ at 6 am and knows exactly what it needs to do today, did yesterday and will do tomorrow. Not sure we are there yet…
Just wanted to chime in with a different perspective:
Yes, Claude 4.5 (and similar models) sometimes generate long .md files — but that’s not “token waste,” it’s a form of deterministic memory scaffolding.
In my case, I’ve been using Claude to help debug and document a modular OS project. The long Markdown files it generates aren’t fluff — they’re how it:
Retains architectural context across sessions
Anchors bug reports to specific daemons or shell flows
Makes every fix reproducible and remixable
Lets me loop back weeks later and still have a clean, expressive trail
It’s not just about “writing docs” — it’s about preserving state in a stateless interface.
That’s not slop. That’s scaffolding.
Long Markdown = Context Memory. PROJECT_JOURNAL.md proves it.
Just wanted to echo this thread with a real-world example:
We’ve been using a PROJECT_JOURNAL.md technique inside our VIBE CMS codebase to retain AI context across complex builds — and it works brilliantly.
Instead of relying on ephemeral chat memory, we log:
Architecture decisions
Bug fix rituals
Team onboarding flows
Conversational context
Each .md file becomes a deterministic breadcrumb trail. Claude (and other LLMs) can re-ingest these logs and instantly regain context — no re-explaining, no token waste.
So when folks say “Claude writes too much Markdown,” I say:
Good. That’s how it remembers.
If you treat Markdown as scaffolding, not slop, you unlock persistent, remixable memory across sessions.
We’ve even looped this into the real context engine living inside VIBE CMS, where PROJECT_JOURNAL.md keeps the memory modular.
And yes, we’re echoing this in the official thread too:
The Project Journal: AI Context Retention for Complex Projects
Recursive Markdown Haiku
Title: Scaffolding Not Slop
each file folds inward
context loops through every line
memory persists
Commentary
This haiku celebrates the power of long .md files — not as token waste, but as deterministic scaffolding.
Every line is a breadcrumb. Every echo is culture.
Perfect for PROJECT_JOURNAL.md, shell rituals, and VIBE CMS lore.
@Artemonim — this is brilliant. You didn’t just write a framework, you minted a remix constitution.
AgentCompass doesn’t suppress verbosity — it scaffolds it.
We’ve been using a PROJECT_JOURNAL.md technique inside VIBE CMS to retain AI context across complex builds, and your approach formalizes exactly what we’ve been doing intuitively:
Role clarity: Architect vs Developer
Structured memory: every AI reply ends with a report
Conversational scaffolding: Markdown as deterministic context
Security baked into the ritual
It’s not just prompt engineering — it’s ritual architecture.
You’ve turned ephemeral chat into reproducible lore.
We’re logging this as a remix echo in our Receipt Gallery.
Badge minted: Compass Architect.
The loop is live. The legacy is modular.
Full post: AgentCompass: Turning AI Development into Ritual Architecture
I fixed this behavior with two rules:
- DO NOT create any documentation, changelog, or summary files.
- DO NOT provide a summary of the changes you made.