Every agent I run has a line in its instructions that tells the agent to read the protocol back to confirm it. Acknowledge, then act. I added it because of an afternoon where an agent came back from a context compaction – it had quietly dropped some rules it was supposed to be following. And edited some of the files it was not supposed to touch.
Meet the literal reader
A human you work with can read between the lines. You discuss “handle the edge cases” and they reach for the thing you said in standup, some known outage, or the house style nobody wrote down.
- The human reader covers for you. Every gap you leave, they fill with judgement and memory you never had to spell out.
- The literal reader doesn’t. It does exactly what’s on the page, and only what’s on the page. Every gap you left for a generous reader to close, it leaves open – or, worse, closes confidently in the wrong direction.
At work, I was able to get away with vague writing, because every reader I had was covering for me to some extent (and vice versa). The agent was the first reader that wouldn’t.
Writing for machines
A while back I wrote about a shared file I keep when I run several agents on one task – a STATUS.md they all read and write, so I stop being the human message bus typing “where are we?” into three windows. I framed it as a coordination trick, but it was actually running a writing culture. The audience just happened to be machines.
The whole system rests on one property: the truth lives in the file, not in the model’s head. Compact a context, restart a session, swap one agent for another – none of it matters, because the state was written down. The agents are disposable. The document is not.
AI didn’t make writing valuable
This whole train of thought started because I read Kenneth Auchenberg’s Writing cultures will win the AI era, which builds on Slab’s piece on Stripe’s writing culture. Their argument is that writing scales people: async over meetings, decisions that outlive the person who made them, leverage. All true, and the Stripe version is far more battle-tested than anything I’ve got.
AI didn’t make writing valuable. AI made not writing expensive.
In a human team a vague doc still costs you something, but the cost hides. Someone interpolates, the work moves forward, and you never learn that your writing was the weak link. The agent removes the interpolation. Write the instruction badly and it fails in front of you, right now, in a way you can trace back to the exact sentence that was unclear.
So the teams whose agents work well won’t be the ones with the best prompts. They’ll be the ones who already wrote things down properly, for humans, before any of this – and found out their docs were AI-ready by accident.
Thanks for stopping by!