Published on

Document Everything!

Document Everything!

The highest-leverage habit in the AI era is embarrassingly old-fashioned: write it down. An agent can read everything you've ever documented and none of what you haven't. The model has read the whole internet; the only thing it hasn't read is you. Every decision that lives in your head, every convention that lives in the hallway, is invisible to the machine that's supposed to be helping you.

The industry figured this out and renamed it. Karpathy called context engineering "the delicate art and science of filling the context window with just the right information for the next step," and within a year it displaced prompt engineering as the skill that matters. Clever phrasing was never the scarce input — the recipe is public and everyone has the same models. What's scarce is the material only you can supply. And you can only supply what exists in writing.

If you want proof of what written-down context is worth, look at who's selling it. The "second brain" stopped being a metaphor for a tidy notes app and became a product category: Mem rebuilt itself around an AI that acts on your notes, and Limitless hung a pendant around your neck to transcribe your days — until Meta acquired it in December. Beneath the agents, memory became infrastructure with term sheets: Mem0 raised $24M and became the memory layer inside AWS's agent SDK; Zep timestamps facts so "I moved to Tokyo" replaces London instead of joining it. And at the top of the stack, Glean pitches the enterprise a "context graph" of everything your company knows — past $300M in revenue at a $7.2B valuation on essentially one sentence: we know your company, the model doesn't.

Software got there first. Docs switched audiences: CLAUDE.md and agents.md files brief the agent on how the repo works, llms.txt marks out what a crawler should read, and Mintlify reports nearly half of documentation traffic is now AI agents — it serves them markdown while humans get HTML. Docs used to persuade people. Now they brief machines.

Even the platforms are fighting over your written record. Anthropic made memory free for every Claude user and shipped an importer that moves your memories out of ChatGPT; OpenAI answered with Dreaming, which tends your memory in the background while you're away. Companies don't build moving vans for a convenience feature.

Here's what changed, though. Documentation used to be overhead — you wrote for a future reader who might never show up, so mostly you didn't. That excuse is gone. The reader is guaranteed now. It arrives in seconds, reads every word, and acts on what it finds. Writing things down stopped being housekeeping and became programming: every fact you record becomes something a machine can do for you, forever.

So narrate your work. Write the decision down, and the why next to it. Turn the hallway conversation into a paragraph. Undocumented knowledge used to fade; now it never existed. The machines can only amplify what you wrote down — so document everything.