Grab a cup of coffee (or tea, no judgment ☕️) and settle in for a chat about AI agentic coding tools.
This is a personal review (no benchmarks, no automatic evals)... I promise!
Cursor
The good parts
- Model Freedom - The coolest part is that you can point Cursor at any major model. Want a little GPT-5 action? Sure. Prefer Claude or Gemini? Heck yeah! Some models are better for planning (ask mode), while others shine in agent mode.
- Multiple Rules Files - You can configure a folder full of different rules for, and Cursor will remember them. That's super handy, esp. when you're juggling micro-services.
-
Killer UX - The interface feels polished:
- You can review snippets, accept or reject them, and see diffs right inside the file.
- Need to undo? Just click the previous message + enter and poof! Instant revert, no
git
dance. - Conversations are saved automatically with neat, user-friendly titles. Allowing to start right where you left off.
- Drag and drop files from the side pane into the chat area; this makes sharing code with the assistant a breeze.
But... (because there's always a “but”) 😅
- Pricing Uncertainty - unexpected changes in pricing is worrying. Especially they don't own the models, Cursor can be easily affected by model vendors pricing changes.
- Lazy Mode? - Sometimes the agent decides it's done (or done trying) after a handful of attempts. Not sure if this is the model or the system prompt 🤔
Claude Code: the Sorcerer with /init
Claude Code is a command-line wizard. It runs right in your terminal and gives you direct access to Anthropic's models.
The awesome bits
-
/init
is magical - Runclaude init
and it scans your whole project, then spits out aCLAUDE.md
file. This acts as a brain dump of your project's architecture, tools and conventions. See Anthropics docs: https://www.anthropic.com/engineering/claude-code-best-practices -
/context
command - Ever wonder why your agent context is full? This command shows what parts are passed into the context. - Top-notch docs - Anthropic's documentation is clear and thorough!
- Hooks & subagents - Want to set up custom workflows? Think of it as building little AI helpers for specific tasks (I haven't tried this, but it looks promising).
The non-awesome bits
- No built-in code indexing - The model decides which files to touch on its own. That can pull in huge amounts of code and spike your token usage (and bill 💸).
-
One big rules file - All your guidelines has to live in
CLAUDE.md
. - Limited VS Code integration - E.g: you either accept or reject the entire file change. No partial approvals like Cursor.
-
No easy revert - If Claude makes a mess, you're back to
git
to roll it back.
Gemini CLI
Google's Gemini CLI is a tempting alternative: very large context and much cheaper.
What's interesting
- Crazy cheap - Gemini Pro 2.5 is much cheaper compared to Sonnet 4.0
- Great analysis - It's very good at static reasoning tasks, e.g: comparing features vs tests or hunting down version conflicts. I've seen it figure out very subtle bugs.
What's not
- Rule rebel - It tends to ignore project rules unless you keep nudging it.
- Queue UX - The queued-message behavior can be clunky; it doesn't interrupt-and-resume smoothly like Claude.
- Messy output - Often requires linting or manual fixes. E.g: it doesn't automatically run a TS compiler check after generation.
- Manual saves - Conversations doesn't not auto-save; forget to save and you might lose the convo (maybe there is some setting that I missed)
- Poor Docs - One github README is all you got 😄
Final thoughts
So, what do you think? Do you have some other AI coding secret sauce?
Top comments (0)