The slot machine myth: ChatGPT compared AI output quality to a slot machine. I don't find this at all – output is predictable when you obsess about what goes in. The bottleneck shifted from capability to context. Tweet

Cowork vs Claude Code: Not terminal vs GUI – it's whether your instructions, skills and agents build on each other. Blog post

Output styles vs CLAUDE.md: CLAUDE.md appends to the coding system prompt. Output styles replace it. That's the difference between fighting defaults and changing what the tool is. Tweet

Founders who hate the terminal: Four commands to get in. The moment it clicks: second session, Claude already knows your company. Blog post

Content entropy: AI defaults to adding, not subtracting. Every edit quietly expands what you didn't touch. Bloated docs become context for the next session. Tweet

Session discipline: Your sessions are too long. One task, commit, exit, new session. New sessions are free when your CLAUDE.md is solid. Tweet (sessions) · Tweet (compaction)

On my radar

Anthropic's multi-context-window playbook: How to handle long sessions. They recommend fresh windows over compaction in some cases. Docs

The 1M context debate: Bigger context hurts coding quality. More tokens, less focus – 150-300k sweet spot, fresh session after that. The window exists as headroom, not a target.

The AI Vampire: Steve Yegge on whether heavy AI usage costs more energy than it saves. The question many building with AI are quietly asking themselves. Essay