I have almost stopped telling the AI what to do. Even so, it does not turn into dumping the work on it — I get what I hand over ready first, and it is the words that I have cut back on. The series "The Art of Not Telling" is the record of building that state one mechanism at a time, and this article is the finale. The mechanism from each article can be reached from the list in the introduction.
There is one thing I have to confess.
The truth is, I do a lot of telling
I spent 3 series and 37 articles writing "not reading," "not listening," "not telling," and this is what it comes to.
I wrote the evidence up in article 1. I added "always run this before committing" to the instruction file only 2 times, and both were the day I put that hook down. The person who made that hook was, in the same commit, writing the same thing out in words as well. It was not that I forgot and then duplicated it. From the moment I put it down, I was writing it twice.
And for 5 months, not 1 line of the words has gone away.
The sentence I wrote most often in this series is this one.
I thought …. It was not so.
I thought that if the machinery was there, the whole section could go (article 1). I thought there were no procedures left on my machine (article 6). I thought that fixing the number when I handed it over would get the row filled in (article 7). I thought the names and the markers matched 1 to 1 (article 9). I thought they were left hanging in the air (article 10). What the 26 automated tests were not looking at, I did not know until I counted (article 11).
I counted
In the introduction I had you count the words you had typed yourself. I counted mine too. While I was writing the last 8 articles of this series, I typed 11 times. 8 of those were the 2 fixed sentences I had written in the finale of the first series.
The wording, though, was not the same. "Are there any gaps" had turned into "Is there anything missing," and "hand the work off to the next session" into "Go ahead with the next task." I did not remember the fixed sentences I wrote myself, not exactly.
What I wrote in the introduction to the second series came straight back at me.
The AI does not remember a rejection. For you it is the second time; for the AI every time is the first.
I was the same. I do not remember what I put down, so every time, I write it in words as well.
The finale of the first series says this.
Through all of it, I read almost nothing. And I say almost nothing.
The first series took 14 articles to explain the not-reading side. Explaining the not-telling side took 13.
What I stopped doing
| The Art of Not Reading | The Art of Not Listening to the AI's Opinions | The Art of Not Telling |
|---|---|---|
| Not reading test output | Not weighing up proposals | Not asking for verification |
| Not reading memory | Not writing prohibitions | Not giving specific instructions |
| Not reading unit tests | Not creating rules that forbid | Not writing up bug details |
| Not using skills | Not writing automated tests | Not asking for a plan |
| Not reading rules | Turning down the AI's requests to check | Not driving the point home |
| Not tidying shared memory | Not listening to what the AI says | Not interrogating the AI |
| Not reading fix history | Not tidying rules | Not criticizing the AI |
| Not reading handoffs | Not choosing from the options | Not letting the AI think |
| Not cutting in | After the first check, not checking the subagent input | Not answering questions |
| Not reading work results | Not having the AI review | |
| After the first read, not reading the subagent logs | After the first time, not spelling out details to the subagent |
31 of them. Reading, listening, telling: I stopped all of it.
The last 3 are the articles in the imperative, one sitting in the same place in each of the 3 series. Read your subagents' output, and only that (series 1, article 5), check the input, and only that (series 2, article 4), spell out the output, and only that (series 3, article 6). Read it, check it, spell it out. Of the 37 articles, these 3 are the only ones I wrote in the imperative. Each of them was a single time, and that one time is what taught me where the mechanism goes. Once it was down, I stopped reading, I stopped checking, I stopped spelling out. That is why they are in the table.
Lining them up, I noticed something. The mechanisms are all down already.
I had simply forgotten that.
From here on, I no longer have to do anything.
For how to place the mechanism from each article, start from the list in the introduction. What was built with these mechanisms is running at typingtube.
Series: The Art of Not Telling
- ← Previous: 11. The art of not having the AI review
- All articles: Introduction