This is article 11 in the series "The Art of Not Reading." It lays out symptoms that go wrong and their remedies, one at a time. Each article is finished once you put down a single file or script. Why that mechanism is needed becomes clear when you read the explanation afterward. The whole picture and the list of articles are in the introduction.

This time it is about a human dropping a word in while the AI is working. A human's instruction outranks everything else inside the AI — so cutting in partway through the work breaks it, priorities already in flight and all. What breaks, and what becomes of the original work after that, is what this article deals with.

Start by looking for just one thing.

It is the place in your most recent session where you cut in while the AI was working. "That name is off." "While you are at it, fix this too."

Right after that, the response to your remark took over, and a few turns later it had become "how far along was the original work again?" Was that not how it went?

What is happening is a reordering of priorities. A rule loses to an ingrained habit (article 1), and a policy you made it read loses to the task in front of it (article 7). The word you have just dropped in joins that line last. The plan that was running and the TODOs that were stacked up usually lose. The more correct your remark is, the more faithfully and fully the AI turns toward it — and the work breaks, correctly.

On top of that, a human interruption has two weaknesses. One is that a human notices many rounds late. The other is that the word stays in the conversation. A remark dropped into the conversation stays in the history, and it goes on carrying the load we saw in article 2 inside the conversation as well. A human's word is neither cut nor left out: it is read back along with everything else, every time, until that conversation ends.

What I understood fits in one sentence.

A human's instruction outranks everything else inside the AI. Cut in partway through the work and it breaks, priorities already in flight and all. So the remark does not come from a human: a mechanism returns it right after the work.

CC BY 4.0

Why work in flight loses to a word dropped in partway

What makes that word strong is that it always arrives last.

As in article 6, the instruction file and a word typed straight into the conversation both arrive as the same user message. The way they arrive is the same; only the position differs. What that position does is what we saw in article 3: information placed at the beginning and the end of a long input can be retrieved, while the accuracy for what sits between them drops. A word dropped in partway through the work sits at the end. The plan that was running and the TODOs that were stacked up sit between them.

And once a word has landed at the end, it stays there. As the conversation nears its limit, what goes first is, as we saw in article 10, the older tool outputs.

The amount that fits in a conversation is fixed, so the room the word keeps taking up pushes older things out. The plan that was running and the TODOs that were stacked up sit ahead of that word.

The mechanism: there is nothing new to put down

The guidance that stops a bare command (article 1), the error that says the setup is not installed (article 4), the catalog that stops a launch (article 5), the insertion right after a save (article 7), the policy at the moment of a retry (article 8) — the hooks this series has been putting down all stand in for the moments when a human wants to cut in.

And they cut in better than a human does. They land in the same position as a human's word, right after the work: the moment of insertion from article 8. What differs is the content: what a mechanism returns is an error about the very move just made. The AI knows what it is about without being told. It takes it as continuous with what it has just done, fixes it on the spot without knocking its own priorities over, and moves on. It does not get in the AI's way.

So what I notice does not go into the conversation; it piles up in a note. In my own case, while developing typingtube (a web service for practicing typing along with music videos on YouTube), even when I see a broken layout or a name that sits wrong, all I do while the AI is running is write it in the note. At a break I read it back, and it goes into an automated test or a hook (article 6), or into a policy file (article 8). A remark converted into a mechanism arrives from then on right after the problem, every time, in a shape that breaks nothing.

What piles up in the note takes effect from the next time on, and does not arrive in time for the work that is running now.

What I stopped saying

Remarks made during the work. "That part is wrong," "and this too while you are at it": I no longer say these while the AI is running. It is not that what I wanted to say has disappeared. The job of saying it has simply moved from the human to the mechanism. What I type at the start of the work now is one line, "go ahead with the next task," and past that I leave it to the mechanisms.

Next time, the last of the advanced articles, the art of not reading work results. I do not check the work the AI has finished. And the gaps still get found.

Series: The Art of Not Reading