read me if you're a vibe coding "token waster"
the name of the game right now is effective token usage
the name of the game right now seems to be effective token usage
Howdy, ya'll
It seems like anywhere you turn there's a joke/meme/complaint about how these models seem to gobble up tokens like they're going out of style
Partially, it's true - even the new Opus 4.7 now requires more effort (tokens) in order to function properly
But on the other hand, a lot of us still have really poor practices with using these tools that cause major bloating of our context windows
I recorded a short ~20 minute video on some tips for not pwning yourself on token usage in the free group: https://www.skool.com/tech-snack/some-token-optimization-advice
If you have anything to add, let me know!
News Worthy
Opus 4.7 Release
So… Opus 4.7 is out, and it's receiving a lot of mixed opinions.
Here's what I've seen with my testing:
It seems to be a little bit of a regression in terms of actually following instructions.
Two concrete examples come to mind:
When using Obra Superpowers to orchestrate builds, it completely ignores the sub-agent driven executions
Translation → all work done in main thread, millions of tokens wasted because it ignores direct commands
Second example:
Ignores context of questions I literally just answered…
Claude: "should we defer feature X?"
Me: No, we need it
Claude: "okay, ignoring that for now!"
Moral of the story → they're gonna need to make some serious fixes to the harness soon before people start bouncing away
Cursor Acquired by SpaceX
I haven't really touched Cursor since Claude Code became a powerhouse… as someone that doesn't really fine-tune code line-by-line, I never really saw the utility of it.
But apparently Elon Musk does.
They acquired the right to buy Cursor this year for $60bn - WTF?
Shout out to the Cursor team, it was my first ever "daily driver" for vibe coding, I'm happy for them.
You can read about it here:
SpaceXAI and @cursor_ai are now working closely together to create the world's best coding and knowledge work AI. The combination of Cursor's leading product and distribution to expert software engineers with SpaceX's million H100 equivalent Colossus training supercomputer will… 10:11 PM • Apr 21, 2026
Claude Design Mode
Potentially the most controversial, Anthropic released Design Mode for Claude, and I think it's pretty great.
Like most tools, if you simply ask it "build me a beautiful dashboard" or some other vague question… you'll probably get some vanilla stuff.
But I did a video on it recently with a custom-built skill that helps you take PRDs and build designs from there.
Check it out here if you're interested: https://www.youtube.com/watch?v=mJ98eeC0oDg
Things Worth Checking Out
- Karpathy Skills: Someone gobbled up all of Andrej Karpathy's advice on how to use coding agents properly and compiled it into a Skill you can use in Claude: https://github.com/forrestchang/andrej-karpathy-skills
- Claude Fork: I was late to the game on this one, but apparently you can fork conversations now inside of Claude Code. Just type in "/fork" and it will branch your conversation at its current place into a new session. Great when you want to experiment from a certain point without muddying your context window
- Build Your Own X: If you're a vibe-coder that wants to learn how to do stuff better, this open source library compiles a bunch of "build it yourself" versions of popular tools. Maybe you wanna learn how to make a Reddit bot, a CLI tool for your product, or build a browser-based video game. They've got guides for all of it: https://github.com/codecrafters-io/build-your-own-x
- NVIDIA giving away free API access to stuff: NVIDIA is currently offering actual free access to some top open-source models. Grab a key here and plug it in to whatever-you-wish-you-could-do-for-free: http://build.nvidia.com/models
- Compound Engineering Plugin: One Claude Code plugin that's quickly becoming my daily driver for building stuff is the Compound Engineering plugin from Every Inc. I did a video on this one last week, too: https://www.youtube.com/watch?v=PMm_vRk7vfQ
- Stop Wasting Tokens: All those CLI commands that dump into your terminal actually cost a lot for the LLM to read. This library compresses the outputs, saving up to 80% of your token usage! Check it out: https://github.com/rtk-ai/rtk
That's it For This Week!
-Sean
P.S. - I'm getting ready to relaunch my paid community for vibe coding to focus more on going from:
MVP → Production Ready → Acquired Users
We're currently at 63 members, price will go up once it's out, so check it out if you're interested: