More notes on Cursor productivity.

Context is everything. The Cursor agent is now really good at grepping the codebase, evaluating code paths, extending functionality where it finds architectural consistency. It falls down at higher level conceptual things.

One example: I asked the agent to build this blogging platform, but there are two high-level ways to think about serving blog content, Dynamic or Static. Static is better, dynamic is easier. I decided to serve dynamically for now, so I can focus on my prose. Cursor kept getting confused, it would apply patterns that conflated the two approaches. It would correctly find the documentation for both but couldn’t figure out that they were conceptually incompatible.

I fixed it by adding “keep it dynamic, yo” to my prompts and this worked, but it’s clear that this instruction needs to be provided is a more formal way.

Cursor provides .cursorrules, but it’s very high-level, and as you can guess from the lightweight documentation, still pretty basic. The agent will also read from documentation markdown files as it codes, and these can be scoped to different levels of the file tree.

Directories of .cursorrules are popping up, and these are great starting points for fixed architectures, but I’m not convinced that it’s a modality that will scale to complex real-world projects.

Cursor also has Notepads, which serve a slightly different function:

Notepads are powerful context-sharing tools in Cursor that bridge the gap between composers and chat interactions. Think of them as enhanced reference documents that go beyond the capabilities of .cursorrules, allowing you to create reusable contexts for your development workflow.

I suspect the long term vision here is that the AI will analyse the codebase, aggregate the documentation, and generate its own unique per-project rules on the fly. It will then keep these rules updated in response to developer commands.

Ultimately, building and retaining a long-term memory of the domain, finding a scalable solution for self-structuring and self-updating context feels necessary for agents to reach the next level of comprehension.