• john garrish
Today is sponsored by the letter C for Context
Lots of reading today - the context on context
The context development lifecycle: Optimizing context for AI coding agents — Context has become the new bottleneck for AI coding agents, replacing raw code-writing speed. The “Context Development Lifecycle” is how you apply software engineering discipline to how organizations generate, evaluate, and govern knowledge for their agents. Bigger context windows won’t fix this. You need versioned, tested context — the same way you need versioned, tested code.
Company graphs = Context repository — A company’s knowledge graph — entities, relationships, decisions, history — is a great primitive for feeding AI agents. Structured relationships between things that matter. This is the direction that a lot of enterprise AI infrastructure is quietly moving.
Obsidian + Claude Code - 24/7 personal operating system — Greg Isenberg’s setup: write everything in markdown, link it like a brain, feed it to Claude Code. Once “the vault” exists, the agent stops being generic and starts thinking in your voice.
The file system is the new database — Muratcan Koylan built a personal OS for AI agents using nothing but a Git repo — 80+ markdown, YAML, and JSONL files. A lightweight routing file tells the agent which module is relevant; everything else loads only when needed.
The hottest new programming language is English — The original 2023 Karpathy tweet that seeded the “software 3.0” framing. Feels obvious now. It wasn’t obvious then.
Skill Graphs > SKILL.md — Flat skill files aren’t enough. You need graphs that link related concepts so context can flow between them. Agents navigate a domain — rather than just executing a single task in isolation.
The AI PM OS — Aakash Gupta walks through what a PM operating system built on Claude Code actually looks like in practice — Cursor and Claude Desktop connected to automate the routine and free up time for the work that actually matters.
Much of any digital job is now preparing context for AI models. Organizing files in folders, naming everything correctly, introducing things in the right order, and only then asking the AI to do something in clear written English.
I realized: knowledge bases and codebases have a lot in common. they’re both folders of text files with relationships between them, they both have conventions and patterns, and they both benefit from agents that can navigate and operate them
Other goodies
Something big is happening - Matt Shumer — The post that racked up 80M+ views in February.
Baumol’s cost disease — Baumol’s disease is the reason costs keep rising in labor-intensive sectors — healthcare, education, live performance — even without productivity gains. Wages in those sectors track productivity gains elsewhere, but the work itself doesn’t get faster.
“The gains from Claude Code in non-software jobs are gains in software shaped parts of those jobs”
“We’re summoning ghosts, not building animals”
Follows
- Boris Cherny — x.com/bcherny
- Andrej Karpathy — x.com/karpathy
- Lenny — lennysnewsletter.com / x.com/lennysan
- Kent Beck — substack.com/@kentbeck / x.com/KentBeck
- Steve Yegge — steve-yegge.medium.com / x.com/Steve_Yegge
- Aakash Gupta — x.com/aakashgupta
- Yours Truly — x.com/SteelAardvark
Your zen for Weds
Don’t persuade yourself into the restaurant mindset. Slowly over time develop the capacity to see the decision space and move around it.