Archives

Length rules don't fix verbose agent replies. Changing the definition of a reply does – to the Slack message a colleague would send after doing the work. The full instruction file, free to copy.

One skill that holds every skill, agent and CLAUDE.md in my setup to the same standards. It reviewed all 38 of my others on day one – three didn't survive. Free to copy.

An output style fixes Claude's grating default voice in five minutes. Or go further: run a custom agent as the main agent and replace Claude's persona entirely.

It took four builds of the same site to get here. The fourth runs on a reference doc (a contract for every component) and each new page now comes out about 95% right in a single pass.

Six investments turn agents into a software factory. The order you make them in decides whether you get 10x the output or 10x the waste.

Boris Cherny says his job now is writing loops, not prompts. There's a Claude Code command called /loop – assuming that's what he means is the mistake. It's level one of five.

The framework that keeps my AI's customer knowledge trustworthy is data governance under another name – source-of-truth rules, validation constraints, schemas that grow on demand.

Someone else's AI setup looks impossible because it's had longer to grow, not because it's a harder idea than yours.

The fix for graveyard call transcripts isn't a better recorder. It's a pipeline that turns each transcript into agent-ready context the moment it lands.

Granola labels every guest "Them" – useless on 3+ person calls. So I built a replacement in 2 hours.

Bigger context windows fix the wrong half of the AI codebase problem. The half they don't fix gets worse as your codebase grows.

Five files, a session loop and a 30-second ritual. The structured memory system I install on every new project.

Its source code leaked. So I used it to explore itself – 8 parallel agents, 1,884 TypeScript files and a hidden digital pet.

Cowork is the right choice for most knowledge workers – one great assistant, one task at a time. Claude Code is for going past one agent: 10, then 100, then 1,000.

Paste one command and you'll never have to re-explain your company to AI again.