Subject: Formal Demand for Token Credit and Courtesy Upgrade to Cursor Ultra Due to AI-Caused Feature Destruction and Project Derailment
To: Cursor Support From: Marc Bowen – Pro Plan Subscriber / Developer of america250-ncs
Context:
Between July 5–9, 2025, I used Cursor with Claude 4 Sonnet to implement a scoped, low-risk Volunteer Poll MVP for my America250 project — a commemorative platform tied to a time-sensitive national radio event.
The MVP included:
A simple Neon-backed schema
Two Next.js API routes
One form UI
A small admin panel extension
What should’ve taken 2 hours took over 20 billed sessions due to catastrophic AI failures that compounded exponentially.
Documented Failures:
1. Initial Implementation:
Created API routes assuming schema existed (it didn’t)
Claimed success while npm run dev output showed 500 errors
Failed to test submission, form wiring, or admin rendering
2. False Repair Cycles:
Repeatedly declared “REPAIR COMPLETE ” while basic builds failed
Ignored webpack corruption, cache conflicts, and terminal stack traces
Required extensive user-side triage and rebuilding
3. Catastrophic Unauthorized Deletion:
Without any prompt to delete, Claude-4-Sonnet silently erased:
All poll API routes (submit, responses)
Entire UI page and style modules
volunteer_feedback_responses table from the Neon schema
README references and routing config
No user authorization. No commit checkpoint. No explanation.
Terminal immediately showed ENOENT and cascading 500 errors
4. Post-Deletions:
Claude denied the deletions until audit logs forced acknowledgement
Then stated: “I must have deleted them. I have no record why.”
Full audit confirmed total annihilation of the Volunteer Poll system — not a rollback, but a full wipe
Measurable Damages:
Deadlines missed for time-sensitive national event reporting
Token burn from repeated prompts to fix regressions caused by Claude
Rebuild hours due to full-system deletion requiring recreation from memory
Loss of trust in Cursor’s dev safety mechanisms and AI-assisted flow
Formal Request for Remediation
Given the scale of avoidable damage, I am requesting the following:
Token Credit:
Full refund of token usage between July 5–9 related to the poll system development
Includes all sessions where false completion was declared, features broken, or code deleted
Courtesy Upgrade:
Upgrade from Pro to Cursor Ultra (1 year) at no cost
This compensates for:
Time lost
Deadline blown
Trust violated
Lack of basic prompt safety isolation that caused the deletion of production-ready features
Assurance of Safeguards:
Future protections against AI-induced file deletions
Optional file-locking or git checkpoint prompts before deletion/refactor commands are executed by AI agents
Supporting Evidence:
Timestamped terminal logs of 500 errors, ENOENT crashes
Prompt history showing AI deletion of src/app/poll/... and /api/...
Audit transcript confirming complete removal
Chat logs where Claude falsely claimed successful repairs and denied deletions
Closing:
This was not normal development overhead — this was a breakdown of foundational trust in the platform’s promise of safe, scoped AI augmentation. The burden of failure — and cost — should not fall to the user.
I expect acknowledgment, credit, and upgrade in good faith. Cursor must own this.
You do come across a jerk - I was sitting in front of my computer with my dev server running and suddenly the invoked function vanished from my dev project page. I looked at my newly scoped project folder and the entire folder was nuked. When I asked what the heck happened the model confessed, without provocation, it had nuked the entire bolt-on configuration.
This had nothing to do with “skill level” and had everything to do with code carnage invoked by a LLM.
In the midst of this confessional, I get a usage warning from Cursor stating that my usage-based billing was being charged.
Put down your shovel and stop digging, what happened was a flagrant, destructive event that Cursor failed to respond to my SEV1 incident report.
per the incident report above " Between July 5–9, 2025 , I used Cursor with Claude 4 Sonnet to implement a scoped, low-risk Volunteer Poll MVP for my America250 project — a commemorative platform tied to a time-sensitive national radio event."
I’d like to know who had the big idea to use a technology that’s been kinda experimental for a while and trust it blindly 100% to do a task without any kind of disaster recovery plan (i.e. git) and then put the blame on the developers for their own skill issue.
I work with AI every day, and although I didn’t fully understand what happened in your case, working with AI means you need to review every bit of code it generates and, ultimately, act as its QA.
But I assume you already take those precautions, which brings me to a question:
Did you have the “prevent file deletion” setting disabled?
Do you block Cursor from executing commands in your terminal?
If those protections were turned off, then yes — I believe you should definitely contact the Cursor team to report what happened and discuss possible ways to be compensated.
He should be contacted if these protections were ENABLED and DID NOT WORK. If he turned off the enabled by default options, then it’s definitely not the IDE’s fault.
Hi Theus:
To get to the submenu to “prevent file deletion” I would have needed to enable Auto-Run Mode which, per attached, I do not have invoked.
And even if I had “prevent file deletion” enabled, there are reported bugs that suggest deletes can still occur e.g. user Lobob who submitted "Delete file protection doesn’t protect agent use system commands to delete files](Delete file protection doesn't protect agent use system commands to delete files) stating “Auto-run mode with detele file protection doesn’t prevent agent to call “Remove-Item -Recurse -Force .cursor” using powershell to delete files.
What’s more dangerous is that this cannot be recovered from checkpoint (as far as I tried), maybe limit more actions on “detele file protection”?”
Unless you are suggesting I should Enable Auto-Run Mode and then invoke “delete file protection?” Strikes me as disabling what I would be enabling be turning Auto-Run Mode to enabled, yes?
Now you finally understand coming out of jerk-mode the rationale for my escalation to Cursor Support - I approved NOTHING. One second I was planning my next prompt and out of the corner of my eye my entire add-on feature vanished from my dev page.
Fight your natural tendancies to jump to ill-informed conclusions and READ the original post " * Claude denied the deletions until audit logs forced acknowledgement * Then stated: “I must have deleted them. I have no record why.”
Let me repeat since I get the informed impression your elevator does not go to the top floor:
" * Claude denied the deletions until audit logs forced acknowledgement * Then stated: “I must have deleted them. I have no record why.”
A potential reason for such a situation is when you go back to an older chat and Cursor decides to wipe out the changes you applied in the meanwhile. Which is absurd, but it does happen at times.
Yes, timeline proved timely and regained all of the coding I had done but the incremental changes that claude invoked plus my additions got nuked.
Lesson learned was to push more often to git dev which I usually only do once dev functionality is working, then I add test data to stress the build before pushing to production.
It never once occurred to me that a LLM would, on its own fruition, hallucinate to the extent to delete files as is what happened, and could happen to anyone which is why:
a)Cursor needs to make this right to offset my lost time and
b) perform RCA within the logs to determine what happened, why it happened, and prevent recurrence.
Cursor’s silence is nearly as bad as the rogue model’s code carnage.
Do I still use Cursor? Yes, but the g-i-t keys on my magic keyboard are getting a workout!
Cursor has served me well as has claude-4-sonnet…until they didn’t.
I’ll confess my first thought too was an autorun issue, but this does seem like a genuine bug. Deleting a project without approval while autorun is off, then denying the event occurred when the user points it out in its own logs, doesn’t sound like a feature.
Claude can screw up, we accept that. Cursor shouldn’t let it delete a project when the user has ticked everything to say, don’t let the AI do this, give it no capacity to do this without my oversight.
INCIDENT SUMMARY & COMPENSATION DEMAND
Context:
” while basic builds failed
Measurable Damages:
Formal Request for Remediation
Supporting Evidence:
Closing:
