Cursor is losing its edge for my heavy coding workflows

Cursor, the AI-powered IDE, is losing its edge for heavy coding workflows due to degraded context handling, causing hallucinations and disconnected responses, according to a developer's account. The issue persists across models like Claude 3.5 Sonnet and GPT-4o, pointing to IDE-level orchestration problems. The developer is considering alternatives such as VS Code with plugins, custom LLM agents, or CLI tools like Claude Code for more reliable context control.

Cursor is losing its edge for my heavy coding workflows Cursor /en/tags/cursor/ this week that I didn't see coming, and it's making me rethink my entire AI-driven development setup. For the last few months, I’ve been treating this IDE as my primary partner, but the recent degradation in how it handles complex context has turned my productivity into a constant battle against hallucinations. It isn't just a one-off glitch; it feels like the underlying orchestration of how the model pulls in codebase context is starting to fray. When I'm working on a large-scale deployment or refactoring a deep directory structure, I expect the LLM agent to actually "understand" the relationship between my types and my implementation. Instead, I'm increasingly getting responses that look confident but are fundamentally disconnected from the local files. The specific failure pattern The issue usually manifests during a deep dive into existing logic. I’ll use the codebase indexing feature to ask a question about a specific function's side effects, and Cursor will either: 1. Ignore the local context entirely: It starts giving me generic "best practice" advice for a language rather than looking at my actual code. 2. Hallucinate file paths: It references modules or helper functions that don't exist in my current branch. 3. Lose the thread mid-refactor: I’ll start a multi-step prompt engineering task to migrate a component, and by the third step, it's completely forgotten the constraints I set in the first prompt. I tried to debug this by checking the indexing status and forcing a re-index of the entire workspace. I even went as far as switching between different models—testing Claude /en/tags/claude/ 3.5 Sonnet against GPT-4o—to see if it was a model-specific issue or an IDE-level context injection problem. The result? It's the orchestration. Even with the strongest models, the way the IDE feeds the "relevant" snippets into the prompt window seems to be missing the mark. Is it time to move back to VS Code + Plugins? I’m currently weighing whether to stick with the Cursor ecosystem or revert to a more modular AI workflow. The "all-in-one" convenience is great until it fails, at which point you're stuck in a black box. If you're building a practical tutorial or a complex system from scratch, you need reliability more than you need a fancy UI. I've been experimenting with a few alternative setups to see if I can replicate the magic without the overhead: VS Code with specialized extensions: Using tools that allow for more granular control over what context is sent to the LLM. Custom LLM agents: Building a local script to pipe specific file contents into a chat interface via API, which gives me 100% transparency on what the model actually sees.Moving toward a terminal-centric approach where the interaction is more direct and less "magical." Claude Code /en/tags/claude%20code/ / CLI tools: If you are relying on Cursor for mission-critical production code, keep a very close eye on your context windows. Don't trust the "indexed" status blindly. If the model starts acting like it's never seen your project before, stop the prompt and manually paste the necessary files. It’s a pain, but it's better than chasing ghosts in your codebase. Next GLM 5. → /en/threads/8057/