Recently, I asked an AI to help organize a set of articles about personal AI. A few paragraphs in, the drafts suddenly began discussing family letters.
I sent them back. Why would a stranger arriving from a shared link know what we were working on? They came to read about AI and found themselves in the middle of a product review.
We had prepared an entrance exam for an article nobody had read yet.
The background was simple enough. At Spiro, I am exploring how everyday records might help adult children share updates with their parents. The AI and I had discussed the idea at length. “Family letters” had a perfectly clear meaning in our conversation.
The reader was the only person who had not attended it.
Two insiders wrote an article
This exposes an awkward possibility: an AI can become more familiar with its owner while becoming less useful at communicating with anyone else.
Writing has at least three participants here: the author, the assistant, and the recipient. Familiarity between the first two does not automatically provide context to the third.
We had discussed why we chose the problem, what we had tried, and what our shorthand meant. The AI wrote from that background. I read from the same background. Together, we missed the entry point.
This happens in project updates full of unexplained abbreviations, handover notes referring to “the usual approach,” and family messages that announce a decision involving someone the recipient has never met.
A knowledgeable reader is still not a mind reader.
Who is Alex?
Take a hypothetical sentence: “Alex and I finally agreed on the plan.”
A colleague who attended the meeting needs no introduction. A parent may need “Alex, the person I am building the project with.” A new partner may need to know which plan and why it affects them.
The event has not changed. The missing context has.
Asking the AI to be warmer or more professional will not necessarily solve this. A warm sentence can still leave the reader wondering who Alex is.
This is a useful test of personalization: knowing a hundred things about me, can the assistant select the one thing someone else needs to understand this message?
More knowledge should sometimes enable less explanation, because it allows a better choice of what belongs.
Importance to the author is not enough
Some of the drafts also carefully described our research progress without explaining why the reader should care.
The sentences were cautious. They were also asking strangers to sit through an internal review before receiving an idea they could use.
Precision matters. An experiment should not become a product claim. A hypothetical scene should not pretend to be a customer testimonial. But the reader does not need a project-status briefing every second paragraph.
A useful article gives them a question they can take elsewhere: for example, how much cleanup work did this impressive assistant leave for its owner?
That is portable. The exact sequence of our internal attempts usually is not.
Add the missing sentence, not the entire history
Once a writer notices the gap, the opposite mistake is to explain everything from the company's founding onward. Readers who need one sentence of context do not need an onboarding course.
Some background should not be shared at all. Knowing why a disagreement happened does not mean the assistant may tell the family. A partner's confidential information may explain a delay without belonging in the update.
The editorial task is precise: supply what makes the message understandable and withhold what does not belong there.
I now try to read a draft while temporarily setting aside what I already know. Has this person been introduced? Why does this event matter here? What can someone carry away after reading?
That is more actionable than asking an AI to “make it simpler.” A reader has already offered their attention by opening the page. We owe them a way in.