Hello again,
This is a continuation of series of articles about vibe coding and AI agents in everyday tasks.
Cursor AI uses Anthropic’s claude-sonet models that have the highest SWE score and the latest ones are comparable only to gemini-2.5-pro in my view. And honestly Google’s models despite the close numbers look weaker than Anthropic’s.
So a week ago I decided to switch directly to Claude instead of using the re-seller, also I removed taskmaster (task planning MCP) as it turned out it is usually messy in creating task lists, doesn’t clearly understand a context and overcomplicate things a lot therefore I got to a conclusion that its better to have a simple features list but with a holistic view of an architecture/patterns and a mandatory tech stack with indicated language and framework versions to avoid AI confusion at writing a code. I noticed that until I asked to rewrite my NextJS/React app to api-routes and context-provider-reducer patterns the AI was messy, made a lot of bugs and couldn’t come up with a decent code itself. At the beginning I started with a cursor rules found on github and it didn’t help too much as they are basically a copy-paste of each other without a deep understanding of the tech. So I dug myself into React and NextJS manuals and transcribed the gained knowledge into Cursor AI agent compact rules and afterwards in a Claude’s guidance files.
Claude is much more simpler in tuning, several .md files instead of tons of different rules for every occasion in Cursor AI. And imo Claude actually isn’t lacking the rules functionality because of a fantastic feature - Agents. On screenshots you may see my qa-automator agent that meticulously covers by e2e and unit tests my NextJS/React app which I’ve been writing the last month (I will make a dedicated blog post to this later with more details about the project).
I have created several agents in Claude: 𝐅𝐫𝐨𝐧𝐭𝐞𝐧𝐝𝐞𝐫 - with the stack I picked for the front NextJS/React/Typescript/shadcn-ui/Tailwind dev; 𝐁𝐚𝐜𝐤𝐞𝐧𝐝𝐞𝐫 - NextJS/Typescript/Python//Supabase-PostgresSQL; 𝐐𝐀-𝐓𝐞𝐬𝐭-𝐀𝐮𝐭𝐨𝐦𝐚𝐭𝐨𝐫. And of course a cherry on top is - a capitalist, architect, orchestrator, lead of the dream team, just an observer - me, your humble servant.
Overall experience with Claude so far - dope level, addictive, unstoppable. But, just be ready to pay, as it burns tokens as much as you let it to: executing shell commands with 4k+ tokens by reading a terminal output of test runs, or rewriting from scratch something you just asked to move into another directory, etc (you can limit agent tools access btw). It isn’t much but it compounds overtime as your codebase grows.