RubyLLM 2.2 is out! Released it on stage at @sfrubyconf
Coding sandboxes with my new mxc gem.
Skills support with my new skill gem.
In memory, disk, and Active Storage file stores.
Tools can ask a person and wait.
Chats survive a deploy.
Tests without a provider, 23 providers.
All optimized for coding agents.
The release I'm proudest of.
https://t.co/wt3N8S6d4K
I review a lot of code that Claude Code writes, and leaving feedback in a chat box was getting old.
So I built diffreview, a small native macOS app that gives you a GitHub-style PR review for your local changes, and feeds it back to Claude.
It's open source.
I've been building something for my workshop.
Offcut is a furniture design app. Add it to Claude or ChatGPT as a connector and just describe what you want.
It runs in the browser and it's free.
๐ https://t.co/xx6lHavqd1
I've been building something for my workshop.
Offcut is a furniture design app. Add it to Claude or ChatGPT as a connector and just describe what you want.
It runs in the browser and it's free.
๐ https://t.co/xx6lHavqd1
AI agents write good code most of the time. But if you canโt follow a PR in two minutes, you canโt push back on it. You just approve it.
Do that enough and nobody knows the codebase anymore.
What helps me: framework conventions, a plan first, then tests, small changes.
Microsoft says AI agents could become an insider threat. I think theyโre right.
So I treat agents like a new hire on day one. Builders get their own branch and dev keys only. A review agent checks every PR. CI has to pass. A human merges.
Whatโs your setup?
@kartikb753 Yep, that's the core of it. The suite runs as a set, every time, since the model sees all the descriptions together. One difference: I assert on the final answer, not the tool name, so I can rename or split tools without rewriting tests.
How are you testing your MCP tools?
I built an MCP server inside ThinkerNotes and realized Claude was picking the wrong tool.
Which got me thinking about testing these tools since you're basically working with plain text, not code.
This video explains what I came up with. ๐
@olegzanx Interesting, not sure I see why logging everything makes sense for this. If my goal is just to make sure the right tool is selected, doesn't that mean I can ignore any failed attempts as long as it figures it out eventually?
Elm 0.19.3 is out today!
โจ https://t.co/IWxBkP3t87 โจ
Explains the rough roadmap for the next few releases. We are preparing for some exciting improvements and features! ๐
If you love reading/listening to business books or podcasts, but you forget most of it, I've built something for you (and for me as well).
Check out ThinkerNotes, a tool that keeps what you actually took from each book, podcast, and brings it back the moment it's relevant again.
My YouTube channel made over $550k without going viral.
Most of it was contracts from people who watched my videos and decided I knew my stuff.
You don't need big view counts. You need the right few people watching.