After about 7,500 sessions I notice they're all variations on the same move — using cheap LLM tokens to claw back something AI made suddenly much harder: understanding what's in your code, your sessions, your design space. Addy Osmani calls it "comprehension debt"; Simon Willison, "cognitive debt".

I had a couple of days giving Claude Code to engineers who'd never used it. What they reached for first wasn't the chat — it was the tools. Once a tool exists, you don't need to know how to build it to use it. I'll walk through three of these tools, the prompt → skill → tool pipeline that produced them, and the patterns engineers picked up fastest when handed them — plus one uncomfortable finding from the research: AI walkthroughs measurably improve comprehension and inflate confidence about it. Designing for both is the actual craft.