Claude Code's source leaked. The entire codebase was made public so I pointed it at itself (8 parallel agents across 1,884 files). Here's the cache trick worth knowing: child agents share their base setup so running 8 doesn't cost 8x. The codebase has 5 layers of context management (explains why it stays coherent in long conversations) and a hidden digital pet with 18 species and tiny hats. Blog post

I built Techie – a Claude Code plugin that changes who the AI thinks you are. Claude Code assumes you're a developer who wants to write code. Techie assumes you're smart but happen not to be technical – and want to automate your work. Same power underneath but a completely different personality. I'm testing with friends before a proper launch but the repo's live. GitHub

"If agents are doing everything why are you working such long hours?" A friend asked this while testing Techie. Honest answer: I'm conducting and building systems, not doing any of the actual work. I keep asking "how can I never need to think about this thing again?" That's the job now.

Vibe coding risk isn't the UI, it's the data connection. Vibe coding = building apps by telling AI what you want. Lovable landing page? Go for it. But an internal tool querying production is a completely different risk. Engineering should define the API and vibe-coded apps talk through that – never directly to production. Tweet

Homebrew silently moved Claude Code to the stable channel. If brew upgrade recently "downgraded" you, this is why. Fix: brew install --cask claude-code@latest. Tweet

On my radar

Agent memory is outgrowing text files. ChatGPT doesn't know what you told Claude – that's the real cost of using multiple AI tools. People are building proper memory systems to bridge the gap. Someone consolidated context from 13 platforms for $15. Substack

Mixing AI models to cut costs. Running everything through one AI model gets expensive at scale. What's working: cheap models for routine tasks (~90% cost reduction), with the most capable models reserved for complex reasoning. Meanwhile Xiaomi's latest is benchmarking near the best available. Quality is converging faster than pricing. VentureBeat

Cursor retrains its AI coding assistant every 5 hours. Not daily. Every 5 hours. Cursor blog