Have AI State the Essence of Your Feedback, Then Make It a Rule── After the second repeat, have AI split your feedback into what changes and why, write it as one rule, and store it where every session reads it

01The operator asked AI to analyse the feedback itself, not just apply it

In September 2026, the operator read a column that an AI had drafted and sent it back for a rewrite. There were two complaints. The prose was wordy. And it took far too long to reach anything about AI.

The instruction listed how to fix it. Open with the essential question. Explain why it matters, then build the piece in the order of what, where, why and how. Close with three key points and a conclusion. Keep that skeleton hidden from readers. Add three flow diagrams made of mixed shapes. Make the title more concrete.

So far, this was an ordinary send-back. The last sentence was different. The operator asked the AI to put the essence of this very comment into words, to reflect on how to avoid getting the same feedback again, and to keep that reflection on record.

As a result, the feedback did not end with one rewrite. It was written down as the top-priority section of the writing rules that every later column draft must follow. Each episode of this series opens with an AI event and a question in its first two sentences because of that rule.

The question worth asking is whether this move, turning feedback into a rule, is a pattern anyone can repeat.

02A fix ends with one draft; a rule reaches the next one first

Feedback can end in two ways. You can fix the draft in front of you and stop. Or you can turn the feedback into a rule and stop there.

The first is quick. The AI fixes the passage, and that draft improves. But the next draft is born with the same habit. The second takes one extra step. In return, the next draft is written to the rule before anyone has to comment.

Turning feedback into a rule changes who it is addressed to: from this draft to every draft still to come.

03Each repeat of the same feedback costs the reviewer more time

Giving the same feedback again and again is more expensive than it looks. Each round, you read the draft, locate the problem, put it into words, and read the corrected draft once more. Even if the fix is the same size, your time is spent once per round.

There is a second cost. The wording drifts from round to round. Saying "wordy" one day and "too much preamble" the next sends the AI slightly different requests. When the wording keeps changing, the AI treats each comment as a separate order.

Episodes 1 to 8 of this series covered ways to turn a vague sense of unease into words an AI can act on: letter spacing, the order of sections, the reader's level, the meaning of a term. Each one made a single piece of feedback more precise. This episode is about not throwing that precise wording away after one use.

04A rule is one sentence with a condition, an action and a reason

To turn feedback into a rule, split it with two questions. First: what will change in the next draft because of this feedback? Second: why is that the right change? The first answer becomes the content of the rule. The second becomes its reason.

Not every comment needs to become a rule. A factual error that applies to one draft, or a one-time change of taste, can simply be fixed. Two signals say a rule is worth writing. You have given the same feedback twice. Or you can see that it will apply to future drafts as well.

A rule fits a condition, an action and a reason into one sentence. Without the condition, it spreads to cases where it does not fit. Without the reason, nobody can decide what to do in a case the rule does not mention.

05In material review, rules drawn from send-backs pay off on the next piece

Pharmaceutical material review is a place where the same comments come up again and again. How efficacy is phrased, how comparisons are shown, how side effects are handled: many comments recur across materials in a similar form. Once reviewers start handing drafts and checks to AI, the same repetition happens with the AI.

Suppose you have told an AI three times that an adjective in its draft reads as overstated. On the third time, stop and have the AI analyse the comment. What changes in the next draft? Words of degree must stay within the range the supporting data shows. Why? Wording that goes beyond the evidence makes the effect look larger to the reader. Written this way, the next draft follows the rule before any comment is needed.

The industry's quality systems already work this way. ICH Q10, the international guideline on pharmaceutical quality systems, asks companies to derive corrective and preventive actions from investigations of complaints and deviations. It says the investigation should aim to find the root cause. Turning feedback into a rule brings the same order into work with AI.

Even with a rule in place, a person still makes the final call. A draft written to the rules can still contain wording the rules did not foresee. Rules cut the work of review. They do not remove review.

06AI does not carry feedback into the next session, and a rule without a reason breaks in new cases

The same feedback recurs because of how the AI works. Anthropic's documentation for Claude Code says each session starts with a fresh context window. Feedback given in one session does not reach the next unless it has been written down somewhere.

The same documentation names two ways to carry knowledge across sessions. One is an instruction file the user writes. The other is notes Claude writes itself from the user's corrections. Both are loaded at the start of every session. It also says that the more specific and concise the instructions, the more consistently Claude follows them.

Adding the reason has support too. Anthropic's prompting guide says that explaining the motivation behind an instruction helps the model understand the goal. Its example is a rule against ellipses. The version that works better adds the reason: the reply will be read aloud by a text-to-speech engine, which cannot pronounce them. The guide notes that the model can generalise from such an explanation.

Research points the same way. In the 2023 Reflexion study, a language-model agent wrote a short reflection in words after each failure, stored it in memory, and read it on the next attempt. The model itself was not retrained. On a coding benchmark, its first-attempt pass rate reached 91%, against 80% for GPT-4.

07Starting tomorrow: after the second repeat, have AI turn the feedback into a rule and give it a home

There are four steps.

1. Stop at the second time. When you give the same kind of feedback twice, have the AI analyse the feedback before it fixes the draft.
2. Split it with two questions. Ask the AI to write what will change in the next draft, and why that holds.
3. Make it one sentence. Include when, what and why. Read it back, and correct it if it drifts from what you meant.
4. Give it a home. Have the AI add it to an instruction file loaded every session, or to the document that sets your working rules. A rule left only in the conversation will not reach the next one.

Do not let the rules pile up. The same documentation warns that longer instruction files reduce adherence. When several rules say the same thing, have them merged into one. When a rule is not followed, check whether its condition or reason is missing.