Claude Code is down. Me, staring at Xcode like it’s 2019.

If you want to try Fastshot (YC F25) yourself, register for RevenueCat Shipyards hackathon and you get free credits. #Shipyard2026

This inspired me to add an Insights tab to Claudit.

My wife is a speech therapist. She told me kids hate doing breath control exercises at home. So I built an app where kids blow into their phone to play games 🎈⛵🎂 10 minutes. One prompt. Built with Replit ⠕ Mobile. 3 games. Progress tracking. Blow detection via microphone.

Day 2 of building Puff Pals 🎮 Yesterday: built the entire app in 10 minutes with one Replit ⠕ prompt. Today: made it actually good. ⛵ Sailboat now has collectible stars + wobble penalty if you blow too hard (teaches breath CONTROL, not just force) 🎂 Candles require

The hardest part of building a breath detection app isn't detecting breaths. It's ignoring everything else. Day 3: spent the whole session making sure ambient noise = zero game movement. A phone sitting on a desk should do absolutely nothing. Sounds obvious. Took 3 iterations

A 3-year-old won't read your error messages. They'll just see that red means "good" in one game and "bad" in another and get confused. Fixed: consistent color zones, matching intensity meters, unified result screens across every game. Green = good. Red = too hard. Everywhere.

Reflecting on what engineers love about Claude Code, one thing that jumps out is its customizability: hooks, plugins, LSPs, MCPs, skills, effort, custom agents, status lines, output styles, etc. Every engineer uses their tools differently. We built Claude Code from the ground up

Built an AI coaching app that runs 100% on your iPhone. No servers, no accounts, no data leaves your device. Pick a coach. Have a real conversation. Get clear takeaways. Coaches remember you across sessions. That part feels like magic. Built with Apple Foundation Models +

Check your last 10 Claude sessions. How many actually needed Opus? I was using Claude Opus for EVERYTHING. - File reads. - Debugging. - Todo lists. - Task tracking. Then I checked my usage data. 917 reads + 207 greps - all on Opus The rule is simple: 🔵 Haiku → explore,

My biggest takeaways from Boris Cherny: 1. Coding is now “solved” for most use cases. Boris hasn’t written a single line of code by hand since November, with 100% of his work now authored by Claude Code. At the same time, he remains one of the most productive engineers at Anthropic,

Learning to use Claude Code efficiently: Run /insights, /compact, and simple code searches on Haiku. Save Opus for architecture decisions and complex multi-file work. I burned 42% of my session limit running /insights on Opus 😅. Same result on Haiku would cost far less Right

Do AI providers have an incentive to NOT do this automatically? They charge more per token on bigger models. Auto-downgrading users costs them revenue. Data centres are consuming enormous electricity. A lot of it is Opus-class models answering questions Haiku could handle.