Insanely well-deserved. I just keep getting more and more value out of these. Just used /prototype for the first time and blew my mind. 🤯
Thank you, @mattpocockuk !!!
This is so true! I used to have multiple things I’d be working on with Opus, but with Composer, by the time I switch to the other task I was going to work on, it’s already responded on the fist!
Composer 2.5 fixed my ADHD coding.
It's so fast, I don't have time to do something else.
And it's good enough for most tasks, so it's my new default now.
I was using Claude Code for the last several months. I felt like I was getting way more bang for my buck. The Composer 2.5 results, however, were extremely intriguing. So I put $20 down on @cursor_ai just to test Composer 2.5, which is exactly as advertised and more.
What I didn't expect, however, was that while I was away, Cursor team made the DX absolutely elite! I am loving the (new?) Agents window. It's so fast. The hotkeys are exactly what they should be. And Composer 2.5 is slaying it.
I'm back. Well done, Cursor team.
It held the plan very well! There were a couple of things we needed to fix, but those were all related to context Claude had in memory or Claude.md that Cursor didn’t know. It deviated on only one implementation, but Claude later said in the audit it preferred Composer’s implementation to our original plan.
Built a plan in Claude with Opus 4.7 and @mattpocockuk's /grill-with-docs skill, then gave it to @cursor_ai's new Composer 2.5 model (which executed absurdly fast), then went back to Claude and asked Opus 4.7 to audit all uncommitted changes against our intent:
"The work is genuinely high-quality. Typecheck and lint are clean, conventions are mostly respected, and the architecture matches ADRs 0007/0008/0009 closely."
Bullish on the future of xAI and Cursor...
Absolutely loving the speed of Composer 2.5 so far. But the 200K token context window is a little jarring after working with 1M Opus 4.7 for the last several weeks.
Mm. Maybe. I see where you're going, but my experience is that Composer 2.5 consistently hits around 15-25% context after the initial prompt, with tokens overwhelmingly labeled as "Conversation," which I assume is just the agent's thought process.
Additionally, the max context needed is also a function of use case: if I'm using Matt's /grill-with-docs or /improve-codebase-architecture skills, I just don't see a way to use them with Composer 2.5 if I'm sitting at, say, 21% context before the grilling even begins.
Absolutely loving the speed of Composer 2.5 so far. But the 200K token context window is a little jarring after working with 1M Opus 4.7 for the last several weeks.
I am in love with the /improve-codebase-architecture skill, but I found myself wanting to take on more than is wise at once.
So I created docs/architecture-backlog.md and edited the skill to consider this document before its exploration phase as a subtle clue to what we've already considered in the past, why it wasn't the right time to make that change, and weigh whether now is a better time given the new state of the repo.
@mattpocockuk good idea? terrible idea? Does my concern make sense and would you do it differently?