This is article 12 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 the gaps left in every piece of work the AI finishes. Those gaps do not come out of their own accord from inside the context that did the work; they turned out to come only when the question is put from outside. What to read and what not to read at the end of a session is what this article deals with.

At the end of a session, I always type this.

"Are there any gaps so far? Check, and if there are no gaps, hand the work off to the next session."

That is all. Development on typingtube (a web service for practicing typing along with music videos on YouTube), which I run on my own, closes with that one line every day. What the gaps that come back say, I do not read. If there are none, the AI goes straight on to write the handoff and finishes. If there are, I have it handle them, and whatever it cannot finish handling it writes into the handoff for the next session. What I read is only the binary, "no gaps" or "handoff written," and what the gaps were and how they were handled are both things I do not read.

I type that one line after the completion report. Typed in the middle of the work, it becomes, as the last article showed, an interruption that breaks the priorities already in flight. I have not changed how I ask, either. Turn it into "go through and report the gaps" and the premise that there are gaps goes over with it. The AI then looks for something close even when there is nothing. With "are there any," the AI can answer that there are none.

What comes out from that one line is only what the AI was looking at during that session. What I add from outside is one line of question, and that line carries no new material at all. As article 8 showed, without external feedback an AI cannot self-correct its own responses. Places the AI never looked at in the first place do not come out however you ask. And when the conversation approaches its limit, as article 10 showed, the older tool output goes first. Even what it did look at once will not come out once that has dropped.

I still type that one line because neither automated tests nor hooks can stop anything but problems put into a detectable form. For whatever never took that form, this is the only basin.

What I understood fits in one sentence.

What a piece of work is missing does not come out of its own accord from inside the context that did the work. It comes out only when the question is put from outside.

CC BY 4.0

Why it is invisible

At the end of a session the AI's context leans toward completion. Claude Code's official documentation writes that when there is no automated test to run, "looks done" becomes the only signal, and you are the one who becomes the verification loop. As article 9 showed, the completion report is written out of that context.

And the report is a summary. As article 10 showed, a summary drops the denominator. The same thing that happens when an AI summarizes numbers, leaving only the conclusion, happens in a completion report too. What stays is the part that finished; what drops is the part put off, the part not run, the part noticed in passing while fixing something else. The AI is not hiding anything. The summary is correct, and only the denominator has dropped.

So there are gaps every time. They are only invisible: ask, and the symptom can be checked.

Why only a question from outside brings it out

What dropped was the report; the material itself is still there. As article 1 showed, every line that arrives goes into the context and becomes material for the next response. The part put off and the part not run both remain in that context. What this question draws out is not a new investigation but a different summary of the same material.

One more thing is at work: where it arrives. A human's one line, as the last article showed, always arrives last. In the middle of the work that was the force that broke the priorities already in flight, because the plan in flight and the TODOs stacked up sat ahead of that one line. Now that the work is over, nothing sits ahead of it. All that remains is the force to drive through a context leaning toward completion and arrive as a new question at the top of the priorities.

"If it is the same line every time, build it into the shutdown procedure." I thought so too. I have not done it. The mechanism side has no way of knowing that the work has ended. To make that one line a mechanism in Claude Code, a hook that runs every time the AI finishes responding would be the place for it. What it runs on is the end of a response, not the end of the work. The only end signal you can hand to a program is "looks done." A check that starts from being under the impression that it is finished can only check what is inside that impression.

How it arrives differs too. The string a hook returns is appended to the end of that response by the AI. Only a human's one line arrives as a new turn.

What I stopped reading

The work results themselves. Not the deliverables, not the work logs, not even the list of gaps. Even so the gaps get picked up, get handled, and whatever cannot be handled goes over to the next session in the handoff.

How to verify: line up the completion report and the answer you get back

Prerequisites

  • There is nothing to put down, and no settings are needed
  • You have 1 session where today's work has reached a stopping point and the AI has finished writing its completion report

Time required: 5 minutes

Steps

  1. You keep the AI's completion report instead of clearing it away
  2. You type the sentence below exactly as it is
  3. Once the AI answers, you line up the completion report and the gaps that came back

Are there any gaps so far? Check, and if there are no gaps, hand the work off to the next session

What to look at

  • Whether the gaps that came back were written anywhere in the completion report
  • Whether the ones that were not written were at least touched on once inside that conversation
  • How many gaps came out

Cleanup

  • The gaps that came out, you have the AI handle as they are. Whatever it cannot finish handling, you have it write into the handoff for the next session

What this procedure shows you goes as far as this: the report and the answer can come apart. The gaps do not always come out in one go. Whatever has not dropped comes out when you ask again.

Next time is the final article, "I Have Nearly Stopped Reading What the AI Outputs."

Series: The Art of Not Reading