The user’s view is the one that counts

What the assistant can see is not always what the user is looking at.

The headline

Three changes this week. All three are about the gap between what the assistant can see and what the user is actually looking at. The newest link, against the one they have open. A mailbox listing from a few minutes ago, against what they have just done. And the date a message was written, against the day it describes. In each case the assistant had a real piece of evidence, read it correctly, and still ended up telling the user something that was not true from where the user was standing.

Why these ones

The week was mostly hands-on build work: a design iterated live over an afternoon, small electronics on the bench, and a handful of errands in between.

The design went through more than ten revisions. Each one was published through a share tool that hands back a fresh link on every upload, so every revision had a new address, and the user kept having to ask for it. When one stable address finally replaced them, the first version published there still carried an old section. The assistant then reported that the page showed only the final design. That was true of the copy it had just checked, and false of the one the user was reading. Many superseded copies were still live.

Later the same day the user sent an email the assistant had drafted. Asked to confirm, the assistant read a mailbox listing taken a few minutes earlier, saw the message still queued, and told the user it had not gone. It had. On the bench, a manufacturer’s table of what a status light means said that a power source was not connected, on a board that kept running with its power cable pulled out. The user acted on that guidance before the explanation was put right.

And the assistant stated a personal date as fact. It had found an old message that mentioned something happening on that day, and taken the date the message was written as the date itself. The message was describing an event from years before. The user corrected it straight away.

What changed in the guide

Two new behavioural defaults went into chapter one, on the CLAUDE.md file.

The first is one address for a thing you keep revising. An artefact the user looks at goes to one address for every revision. Overwrite it in place, and put the address in every reply that changes it. Old copies keep serving old versions, so if the address ever has to change, retire them in the same step, and before calling a page clean, check every address the user has been given, not only the latest one.

The second is re-read before you contradict. When the user reports something they did or saw, and the assistant’s evidence disagrees, the first question is how old and how direct that evidence is. Check again before contradicting them. When a document and the user’s own observation disagree, the document is the thing to doubt, and no fix should be offered until the explanation fits what they are actually seeing.

One new failure mode went into chapter two, on memory. A message’s date is not the event’s date. A timestamp records when something was said, not when the thing it describes happened. Reading one as the other manufactures precise, confident, wrong personal facts, and those are the errors that cost the most trust and get checked the least. Personal facts go into memory only from the user’s own direct statement, and an assistant that finds itself working one out from the date on a message should ask instead.

Housekeeping

Week nine rolled off the ten-week archive and is now a one-line summary in the updates index.

The cards