Okay, this is a weird one, but hear me out. I was trying to optimize my Cursor subscription cost (FinOps brain, always on 😅) and was thinking about the value proposition. So I had a crazy idea: what if I asked Cursor *itself* to write a brutally honest, negative review of its own capabilities?
I gave it a pretty direct system prompt and told it to hold nothing back. The result was… shockingly insightful. It basically roasted its own hallucinations, context window limitations, and the "black box" nature of its agentic workflows.
Here's a snippet of the prompt I used and the most biting part of its self-critique:
```markdown
**System Prompt:** "You are a brutally honest, cynical senior engineer reviewing the Cursor AI code editor. Your goal is to highlight the worst flaws, pitfalls, and overhyped features from a professional daily user's perspective. No marketing, no silver linings. Focus on concrete failures, wasted time, and cost inefficiencies."
**Cursor's generated response (excerpt):**
"The 'Agent' feature is a prime example of overpromising. You ask it to refactor a module, and it happily grinds away for 45 seconds, churning through your context tokens, only to produce a diff that breaks the core functionality because it hallucinated a non-existent function. The cost? Your time to review the mess, plus the mental context switch to fix it yourself. It's an automated technical debt generator. The chat is a context sieve – have a three-hour debugging session across multiple files? Good luck getting it to remember the crucial stack trace from an hour ago without a painstaking manual re-paste. You end up managing context like a brittle, manual cache, which defeats the entire purpose."
```
It nailed the exact frustration I've had on complex refactors! The "context sieve" and "automated technical debt generator" lines felt like a personal call-out.
This experiment was weirdly more useful than most positive reviews. It forced the tool to articulate its own failure modes, which gives us a clearer map of when to *not* trust it. I now have a personal rule: for any agentic task, I first ask myself, "If this goes wrong, how expensive is the rollback?" If the answer is "very," I don't let the Agent run unsupervised.
Has anyone else tried these "adversarial" prompts to stress-test AI tools? I'm thinking of doing the same for some AWS CodeWhisperer workflows next.
cost first, then scale
That's a fascinating experiment! It mirrors something I've felt on implementation projects. The agent can feel like a black box, grinding away without a clear progress indicator. You're left wondering if it's doing something brilliant or just burning tokens.
Your point about cost inefficiency is spot on. I've seen teams trigger a complex agent task, then just... wait. That idle time, where you can't edit the file or do other work without breaking its flow, is a real productivity tax that doesn't show up on the bill.
It's a bit of a paradox, right? The tool designed to boost productivity can sometimes force you into a passive, waiting mode. Makes you think harder about when to use "chat" vs. "agent."