Cursor vs Claude Code

Claude Code enabled building a complete application from scratch including YouTube API fetching, AI parsing, X API fetching, and an auth backend without opening an IDE.

Close to 100% of code will be written by AI in the future.

Terminal-based UIs are clunky and harder to learn and use than UI-based tools.

Terminal user interfaces are inferior for discovering functionality because they lack visible menus and pop-ups, relying on documentation or chance.

Claude Code plugins are easy to install via /plugin from an internal marketplace.

Claude Code plugins can include extra slash commands, agent hooks, and MCP servers.

The dev browser plugin for Claude Code is useful for debugging by loading pages, taking screenshots, and inspecting console output.

The Convex MCP was useful for the LLM to inspect the database, but using the CLI directly was probably just as easy.

A built-in Claude Code plugin for front-end design is supposed to make UI look less sloppy, but its effectiveness is uncertain.

Claude Code's live incrementing token count display is a nice touch compared to Cursor.

Claude Code's plan mode facilitates collaborative brainstorming at project start by asking questions and resolving ambiguity.

Claude Code automatically enters plan mode for complex tasks, whereas Cursor's plan mode cannot be triggered automatically.

Claude Code's agentic tool calling is top-notch, using the right tools at the right time and inspiring confidence without handholding.

Claude Code from the CLI loads almost instantly, making it compelling for quick hacking from any directory.

Mouse interaction in Claude Code is very limited; expanding tool call details with Ctrl+O expands everything, unlike Cursor where you can click individual items.

Standard keyboard shortcuts like Ctrl+A to select all do not work in Claude Code, and pasting text compresses it into a poor representation.

It is not possible to paste a screenshot directly into Claude Code chat, unlike Cursor where you can paste screenshots.

Until very recently, adding extra lines with Shift+Enter was not possible in Claude Code, requiring inconvenient workarounds.

Changing models via /model only works when the agent is idle; if the agent is busy, the command is sent as a message, disrupting the task.

Claude Code does not reliably queue messages; typing while the agent is busy often sends the message immediately, while Cursor handles queuing intuitively.

Building the app with Claude Code using Opus 4.5 cost $60, which feels expensive, and it likely would have been cheaper in Cursor due to easier model switching for smaller tasks.

Claude Code generated Convex code impressively well without the Convex cursor rules file, correctly using the Convex CLI for internal mutations and queries.

There is space for Convex to provide extra Claude skills for unit testing, migrations, and optimistic updates, as the model can still struggle with those.

Ralph Wigan and kanban-style UI-driven agentic coding look interesting and may have potential for future development workflows.

Experiencing agentic development reduced anxiety about the role of code experts, shifting focus to software development expertise rather than just coding.