Good Enough for Who?

Samuel Pepys wrote his diary in Shelton's Tachygraphy — a commercial shorthand system he learned from a textbook, not an invented cipher. It was perfectly legible to anyone who knew the system. The problem was that by 1800, Shelton's Tachygraphy had fallen out of use. When a student named John Smith was hired to decode the diaries at Magdalene College, he spent years working through the symbols without a key. The decoding guide — Shelton's original textbook — was sitting on the same shelf as the diaries. Smith didn't know to look there. He eventually figured it out, but only after solving it the hard way.

Pepys' notes were good enough. They were good enough for Pepys, which was the only standard that mattered to him. He was not optimizing for future readers. He was writing quickly, efficiently, in a notation system that reduced effort at capture time and recovered perfectly at recall time — as long as the reader was him.

Most note-takers operate on the same implicit contract.

The threshold you don't notice

Herbert Simon introduced the concept of satisficing in 1956: instead of searching for the optimal solution, people search until they find one that clears a threshold they've set in advance. It's not laziness. It's a reasonable adaptation to a world with limited time and limited attention. You satisfice when you choose a restaurant that's fine rather than spending four hours finding the perfect one.

Note-taking is satisficing by default. The threshold you set at capture time is calibrated for you, in the moment. "Talk to Marcus about the Q3 thing" clears that threshold easily — you know exactly who Marcus is, you know which Q3 thing, and you know what talking to him about it means. The note costs two seconds to write and recovers perfectly when you're you, reading it an hour later.

The threshold shifts the moment the reader changes.

A colleague trying to cover for you doesn't know Marcus. The version of you who returns from a month-long break doesn't remember which Q3 thing. And an AI agent reading your knowledge base has no working memory at all — no ambient context, no pending intentions, no shared history. For the AI, "Talk to Marcus about the Q3 thing" is three tokens with no referents. It will try to infer. It might sound plausible. It probably won't be right.

What the new reader costs you

This is the piece that most note-taking advice gets backward. The standard recommendation is to write more complete notes: more context, more explanation, more legible prose. The problem is that the satisficing threshold exists for good reasons. Slowing down capture to optimize for future readers creates friction that kills the capture habit entirely. A note that's never written because it felt like too much work is strictly worse than a terse note that partially fails.

The better frame is not "write more" but "resolve the references."

The difference between "Talk to Marcus about the Q3 thing" and "Marcus Reinholt, Q3 budget approval for Helios project" is not volume of prose. It's that the second version binds the vague references to durable, queryable things. Marcus becomes a named person. Q3 thing becomes a project and a topic. The satisficing threshold doesn't need to change — you're still writing quickly, still satisficing. You're just anchoring the shorthand to something that outlasts the context you were in when you wrote it.

This is exactly what typed entities do in a knowledge base. A person record for Marcus Reinholt doesn't require better notes about Marcus. It requires that when Marcus appears in a note, you link him to the record that already exists. The reference resolves on the first capture and stays resolved forever. Every subsequent note that mentions Marcus inherits the resolution without extra effort.

The reader you didn't plan for

Pepys died in 1703. He had almost certainly never imagined anyone needing to decode his daily shorthand notes. He wasn't careless about posterity — he donated his library to Magdalene College with meticulous instructions. He just didn't think his operational notes were the kind of thing a future reader would need to navigate. They were written for him, in the present tense, with perfect recall of every reference he used.

Most people's knowledge notes are the same. They're operational — written for the decision or task at hand. The chance that a future reader, human or AI, would need to navigate them feels abstract and distant. Until it doesn't.

The moment you connect an AI to your knowledge base, the future reader arrives. Not at some theoretical point, but every time you ask a question. The AI reads the archive you built for yourself and attempts to use it as if it were a reference document. Wherever you left a reference unresolved, the AI either fails or guesses — filling the gap with a confident inference that may or may not match what you meant.

Typed entities — person records, project records, decision records — are not a formatting preference. They're the difference between a note that resolves and a note that guesses. The reference, bound at capture time to something durable, becomes something any reader can use. The reference left floating in working memory, written in personal shorthand, does what Shelton's Tachygraphy did after 1800: it works perfectly until the person who could decode it is no longer in the room.

Pepys' full diary was finally published in 1825, transcribed by John Smith. It's now considered one of the most important firsthand accounts of seventeenth-century English life. Pepys would have found the whole exercise baffling. He was writing for one reader, in the present tense, with total confidence that the context would always be available.

Most people still are.

---

Asgeir Albretsen is the founder of Harbor.