Using AI to code real complex projects you’ll feel it hit distribution shift when your use case isn’t well represented in the training data.

+Every thread needs to relearn your architecture in a context efficient way you need to be careful that the AI is following patterns consistently.

This feels too melodramatic to me right now.

I use Claude Code and it still messes up the code quite often.

I stopped using it for the unique or tricky parts of our code.

If you’re writing basic React+Tailwind components then it’s great but performance isn’t consistent across all domains.

I haven’t tried out MCP Hub yet, but it looks great, thanks for sharing!

github.com/ravitemer/mc...

This is awesome! FYI it looks like there might be a bug in the new go.bsky.app page.

The x and y axis cut offs are doing a lot of work here too.

It’s just reversion to the mean.

I’d rather just record and transcribe the meeting, then drop the transcript into the LLM for a summary, Q&A or whatever multiple times as needed

Lol, no. The primary reason is definitely phishing.

It prevents phishing attacks. Very common.

You can look up how it works yourself, but Bluesky isn’t doing anything out of the ordinary here.

Not everything is a conspiracy.

www.reddit.com/r/web_design...

This is a generally accepted security practice to protect against malicious links. Many large websites do this.

web.dev/articles/bui...

If you’re just starting out with web development I’d highly recommend checking out other articles on web.dev. It’s a wonderful resource!