The audience you forgot

The note reads: Follow up with her about the timeline. She was worried.

You don't remember writing it. You definitely don't remember who "she" is, what project "the timeline" refers to, or what specifically worried her. The note was made three months ago, during a week that probably had twelve such conversations. The meaning has fully evaporated.

Now imagine your AI reading the same note to draft a follow-up email. The AI has even less to work with than you do.

This is the author-reader gap, and it's as old as writing itself. Notes are built on what linguists call "common ground" — the shared context, vocabulary, and assumptions that make communication compress. When you write her, you and your immediate-future self share the common ground that makes the reference work. Her has a referent. The timeline is one specific thing. The worry has a shape.

That common ground erodes. Faster than you expect.

Herbert Clark and Deanna Wilkes-Gibbs demonstrated in a 1986 experiment how aggressively shared references compress once established. Pairs of people coordinating on abstract tangram shapes developed private shorthand names — "the upside-down martini glass" becoming just "martini" — that worked perfectly between those two people and were completely opaque to anyone else. Your notes are full of this kind of compression. It felt lossless when you wrote it. It wasn't.

The audience is expanding

The problem with private shorthand is that it's becoming less private.

AI agents read your knowledge base now. When something like Harbor connects to your notes, there's a non-human reader querying your documents on demand. When it reads follow up with her, it either guesses, asks, or produces something generic. None of those options existed at write time because you assumed a single reader who already knew the context.

Future you has a similar problem, with more excuses. The version of you returning to a project after three months is functionally a different person with respect to the specific frame you were working in when you wrote the note. You can usually reconstruct enough. But reconstruction is work, and the more compressed the reference, the more work required.

Colleagues who read shared documents face it immediately. A meeting note that says agreed to push back on the original spec is a perfectly useful record for the people in the room. For everyone else, it's a reference to something that existed only in the room.

None of this was a design flaw. It was the correct assumption for the only reader that ever existed: you, soon after writing. What changed is the assumption.

What the engineers noticed first

Software engineering ran into this problem and built a partial fix.

When Michael Nygard wrote about Architecture Decision Records in November 2011, he wasn't describing a filing system. He was describing a form designed for the reader who wasn't in the room. An ADR has a fixed structure: the context that prompted the decision, the options considered, the choice made, the consequences that followed. Someone who joined the team two years later can read it and understand not just what was decided but why — and what was ruled out.

RFC culture in open-source engineering operates on the same principle. A good RFC assumes the reader has none of the specific situational context. Every assumption is surfaced. The context section is written for the least-informed person who might plausibly need to understand it.

These conventions emerged from decades of expensive miscommunication. Decisions without documented context get re-litigated. Systems get rebuilt because no one could reconstruct why the previous design made that choice. The fix wasn't better tooling. It was writing for the stranger who would eventually show up.

What structure actually solves

Typed knowledge entities are partly an answer to this, though they're rarely framed that way.

A person record that links her to a name, a company, an open item, and how you know each other — that makes the reference resolvable for any reader. The private shorthand survives the erosion of common ground because it's attached to something durable. A decision record with context and alternatives fields does the same thing for we went with option B.

The schema is the shared vocabulary that common ground would otherwise need to supply. Any reader who understands what a person record is can navigate your knowledge without having been there. That's a different kind of value than better search or faster retrieval. It's the difference between an archive that only you can interpret and one that works for readers you haven't met yet — including the AI that reads it on Tuesday when you ask it to help draft an email.

Structure is usually justified as a retrieval improvement. The more honest argument is that it changes who can read your notes at all.

Something I haven't worked out

There's a stranger version of this I'm still thinking through.

When you know an AI will read your notes, you write differently. More completely. Fewer private references. You explain things you'd previously left implicit because they were obvious to you. The invisible audience creates pressure toward legibility — the same pressure a public blog creates, where you can't be sloppy in ways you could get away with alone.

Samuel Pepys kept his famous diary in a personal shorthand he believed made him invisible to readers. He wrote honestly because he thought no one was watching. The moment a reader exists — any reader, even a non-human one — the writing changes.

I'm not sure that's good. Cleaner notes are more useful. But the half-formed thoughts, the things written because you needed to write them rather than because they were ready — those might start disappearing. The unfinished thought has its own value.

The audience for your notes just got larger and stranger. And it turns out that changes not just how you organize things, but what feels worth writing down at all.

---

Asgeir Albretsen is the founder of Harbor.