Summary
Once the session ends, an AI agent forgets the feedback it has received, so when it comes to habits of writing or of judgment, the human side ends up repeating the same feedback over and over. To stop that repetition, this site is run on a mechanism where my feedback is written into the operating-rules documents as new rules, and the AI in the next session works according to them. In this article I actually count the operating records of the last two days of running this mechanism, and I write about where in an AI to put my own habits and values, and how to write them so that they take hold, and what comes out of that as a result.
Let me put the counted results first. My feedback remains as dated sections in a document called the record of instructions. When I counted the sections for the two days of July 10 and 11, 2026, at the point I began writing this article, there were 48. Of those, the feedback about how to write text was added, over 17 commits, to a file called the style skill that gathers the writing rules in one place, and at the same point that file held my own words, verbatim, as dated quotations in 11 places. The rules do not end once they are written; they became review points for the pre-publication check. Among the pre-publication checks, the ones that cannot be judged without reading the context are left to an AI, and I call this role the judge. When an already published article that dealt with search strategy was rewritten, this judge caught 6 places that needed fixing, and all of them were corrected. What is interesting is that my feedback becoming a rule, that rule becoming a pre-publication check, and the check catching the fix in the next article, all went around once within the same day I gave the feedback.
There were also failures where things did not take hold as taught. When I taught the value of writing readably, the AI translated it into numbers: an upper limit on the count of commas and on the length of a single sentence. It added that to the rules, but the rewrite that fit those numbers cut the sentences up too much, and the result was monotonous. Translating readability into numbers made the writing harder to read instead. I pointed out that this approach itself was an anti-pattern, threw away the numeric limits, and replaced them with a review that reads the draft aloud and looks at its structure.
In an earlier article, "Two projects, different in field and in build, had sorted where feedback to an AI should live into the very same three layers," I dealt with the criterion for which layer to put a single piece of feedback in. This article is a continuation of that, the installment that deals with the whole act of conveying, that is, what I teach, where it takes hold, and what comes out of it. The body you can read with a subscription begins by showing, with the real figures from the two days, the premise that feedback disappears and the flow that carries feedback from the record into the rules. Next, I divide what is taught into four: the principles of values, the rules of writing, the yardstick for judgment, and the line you must not cross. I write with real examples that each takes hold in a different place. A written rule works only once it becomes a check, so I confirm that with the breakdown of the feedback the judge caught, and I write about the lesson in teaching that I got from the failure of numeric limits, and the caveat that whether things take hold cannot yet be measured, and I close with what came out of this work of conveying.
Audience and takeaways
This article is for people who want an AI agent to work in their own way, whether in coding or in writing, yet find themselves repeating the same instructions every time. You can take away the flow that carries a piece of feedback past the on-the-spot fix and into the rules. I also write about how to change the way you teach and where you place it depending on whether what you teach is a value or a procedure. A rule works only once the AI reads it, so I also cover how to connect the rules to a check and bring them into a form that does not depend on whether they are read. This article is a record of practice based on this site's own operating records.
The gist of those two days
I dealt with this site's own operating records for the two days of July 10 and 11, 2026. By the count at the point I began writing, 48 sections of my feedback remained in the record of instructions, and of those, the feedback about how to write text was reflected into the style skill over 17 commits. The reflected rules became review points for the pre-publication judge and forbidden patterns checked by machine, and they were applied to the rewriting of an already published article. The materials are the record of instructions, the git commit history, the audit records of the operating data, and the record of the pre-publication check of a published article.
Carrying feedback from the record into the rules
To avoid repeating the same feedback, the way this site is run sets the path that feedback travels. When I give a piece of feedback, it first remains in the dated record of instructions. Next, if that feedback is general enough to work on tomorrow's session or on other articles, it is added that same day, as a rule, into the operating-rules documents or the style skill. In this operation, this adding is called burning in. When a rule is burned in, my words go with it as a dated quotation, so that the origin of the rule can be traced later. Finally, if there is a published article that the rule can fix, it is fixed on the spot.
When I counted, at the same point, how much this path was used over the two days, my feedback in the record of instructions came to 32 sections on July 10 and 16 on the 11th, 48 in all. The commits that went into the style skill were 7 on the 10th and 10 on the 11th, 17 in all. There were many cases where burning in and fixing an article happened at once in a single commit. For example, my feedback on the 11th was "do not make the reader read between the lines, and do not write words that make them imagine the intent." That same day it became the rule of restating things in concrete words rather than settling for a suggestive metaphor, and in the very commit that carried the rule, 7 paragraphs of the free part and 16 places in the paid body of a published article that dealt with a mechanism for machine-scoring answers were rewritten into concrete words. The feedback "do not arbitrarily drop what number this is or what cost this is" also entered the rules in the form of always saying, at first mention, what a noun like a number or a cost refers to, and it went on from there to fix an article on search strategy straight through from the top.
A human wakes to an alarm, acts according to ingrained habits and values, and works while watching the changes in the world. I think about this operation with the metaphor of an alarm: for the AI, the schedule is the alarm, the rules burned into the operating documents are the habits and values, and the diffs in the data from my development activity are the changes in the world. So how much of the habits-and-values side I can put into words and accumulate becomes the very substance of this mechanism.
What you teach changes where it goes
In the earlier article I sorted where feedback lands into three layers: the rules documents that are read every session, the skills that gather up fixed procedures, and the memory that keeps the history of decisions. As the two days of feedback were burned in, it became clear that what is taught also comes in different kinds, and that each kind settles in a different place.
The first is values, that is, the principles that set the direction of judgment. My feedback, that plain words are never insufficient and that however advanced the content, the writing should simply be plain, was placed at the head of the readability section of the style skill, as the topmost principle standing above the individual rules of writing. What it teaches is the whole direction of that judgment: when you feel like choosing a difficult word, that is a sign that the thought has not yet been unraveled into words, and what should be fixed is not the vocabulary but the explanation.
The second is habits, that is, the rules for when you actually write text. Open by sharing the background and the premises. At first mention of a noun that needs a referent, such as a number or a cost, always say what it refers to. Do not omit subjects and objects too much. Introduce an abbreviation in parentheses only when you actually use that abbreviation in a later sentence. Such rules lined up as 17 items in a single section, the principles of readability.
The third is the yardstick for judgment. It teaches a criterion you can judge by yourself when in doubt, in place of a list of banned individual words. Whether it is fine to use a difficult word is judged by whether you would say that word in a conversation with a colleague. When this yardstick is in place, the same judgment reaches past the replacements already on the list and works on words that have not yet been listed. Those replacements were all Japanese words opened from a stiff form into a plain one: 傍証 (corroboration) into 裏づけ (backing), 蓋然性 (probability) into 見込み (likelihood), and 寄与する (contributes) into 効く (works). Whether an article is thin is judged by whether nothing remains once you remove the specifics that only someone who has practiced could write, and that yardstick has the same shape.
The fourth is the line you must not cross. The declaration that writing which merely looks the part, with no concrete practice behind it, can never go out under my name did not fit as one more of the writing rules; it was burned into the pre-publication safety gate as a failing condition for publication itself. The line alone is placed where it has the power to stop publication.
The principle settled at the head of the readability section, the rules as items, the yardstick as the criterion for judgment, and the line at the publication gate, each in its own place. Even in the same work of teaching, depending on what is taught, where it takes hold differs this much.
A rule works only once it becomes a check
If a rule is only written into a document, whether it takes hold depends on whether the AI reads it. As I wrote in the earlier article, depending on being read is a weakness, so the rules are converted into checks at the same time they are written. At the end of the style skill, 8 review points to hand to the judge are lined up, and what is checked by machine is narrowed to just 4 items: the bans on words and patterns.
There is a record of this check working. When the article on search strategy was rewritten on July 11 into a single piece of writing, the judge that checks against the readability rules caught 6 in all: 2 introductions of an abbreviation that is not used, 1 chaining of noun phrases, 1 omission of an object, 1 long enumeration that makes you wait for the predicate, and 1 stiff word. All of them were corrected, and every one of them was caught by a rule that had just been burned in. The round trip of feedback becoming a rule, the rule becoming a check, and the check catching the next fix closed within a single day.
This article itself is under that check too. It is written in line with the 17 readability principles and the yardstick for judgment, and before publication it passes the same judge's check.
A value translated into numbers
Not all of it went well. When I taught the value of writing readably, the AI translated it into numeric upper limits, how many commas per sentence and how many characters per sentence, added that to the rules, and even built a check that measures articles by those numbers. As a check it did work, and there is a record of putting the 2 articles I had chosen earlier as models of writing through two checks, one for readability and one for facts, and then fixing the 3 points they raised. But the rewrite that fit the numeric limits cut the sentences too much, and I pointed out that I do not cut my sentences this much, and that setting static numbers for commas and sentence length is an anti-pattern. That is because if the structure is plain and reads from front to back, a longer sentence is easier to read in one go. The numeric limits and the numeric check were abolished that same day and replaced with a review that reads the draft aloud and looks at its structure.
What this failure taught me is where to draw the line on what may be dropped into a machine check. Only the bans on words and patterns, where the call is mechanically black and white, are dropped into the machine check, while values are held in a form that leaves room for judgment: the topmost principle and the yardstick for judgment. This line-drawing itself became one more of the rules burned in over the two days' round trip.
There is still no yardstick to measure whether things have taken hold. In the earlier article too I wrote that whether recurrences of the same kind of feedback had decreased could not be measured, and that is still the same now. What can be counted now goes only as far as how many rules were burned in and the count of feedback the check caught, and whether the number of times I give the same feedback twice has decreased still cannot be counted. Measuring, from the operating records, whether the rules were actually kept is stacked up as a proposal for the work to write next.
What came out of the work of conveying
As a result of these two days of the work of conveying, what I had in hand at the point I began writing this article was the style skill that holds my own words as quotations in 11 places, the 8 review points of the judge that turned them into a check, and the group of published articles rewritten by those rules. The articles already published on this site, up to before this one, were 9 in all, 4 free, 4 paid, and 1 subscription extra, and this article goes out as the 10th to pass the check of the rules burned in over the two days.
A human wakes to an alarm and works according to habits and values. To have an AI work in the same shape, I had to put the habits and values into words, decide where to place them, and go as far as turning them into a check. The work of conveying ingrained habits and values to an AI, and what kind of content comes out of it as a result. I think this may be the first interesting undertaking of this mechanism. Knowing whether there is really value in a world where an AI keeps making content just from my own data being updated is, honestly, frightening too, but since I am a researcher, I want to try making one answer to it.
Originally published at The Future of Humans, AI, and the Web, a site where my research and development is recorded and analyzed by a human and an AI.
Top comments (0)