What's up?

When I start a new project, I typically start by creating some sort of document that describes what I want to build. One name for this kind of document is a Product Requirements Document (PRD) I create the PRD by having Cursor ask me bunch of questions (prompt in comments) and then Cursor creates a PRD based on that Then I review the PRD and ask Cursor to make adjustments if necessary

When building a landing page, don't just ask Cursor to build you a landing page Instead, know that a landing page has several distinct sections and then ask Cursor to build each section one at a time The picture shows the typical sections in a landing page, you can just copy the sections from there Image source

Reminder: Frequently refactor your code Over time, AI gets your code to be a mess. So you should regularly ask it to clean it up Instead of asking Agent to "refactor the code", ask it to: - Eliminate dead code: “Delete unused imports, variables, and functions; point out any test files that still call them.” - Tighten dependencies: "List packages that are imported but never referenced, and suggest how to remove them from requirements." - Kill duplication: “Detect code blocks duplicated ⩾ 3 times across modules; create a shared utility and update call sites.” - Strengthen typing: “Add or refine type hints where ‘Any’ or missing annotations are found; flag places where typing uncovered a bug.” - Demystify magic numbers: “Replace literals that look like config (timeouts, thresholds, etc.) with named constants in settings.py.” - Improve architecture: "Identify features or sections of the app that could be extracted to a separate module." You can also use these to target a specific part of the application Note: the list is not exhaustive! Let me know if you want more material or a course about this

I often have multiple different chats open in Cursor at the same time, one chat for one feature. But why? I do this is to manage the context more effectively. Let's break this down: 1. Context management 2. Context pollution With one feature for one chat, it is much easier for me to pull in the required context and give the AI everything it needs to know. This way I can also roughly keep track of what the model is likely going to modify. Well why not just pull in the whole code base? I call this "context pollution". Too much context, especially irrelevant context, tend to trip up the models. They might start editing the wrong place or make wider edits than they should. On top of that, the model performance and intelligence degrades when the context fills up, and the codebase might be too big to even fit into the context.