This is article 2 in the series "The Art of Not Listening to the AI's Opinions." 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 how you write prohibitions. A prohibition is a rule about what you have decided not to allow while developing. "We do not build this feature," "we do not add this column": decisions you hold both the AI and yourself to. Prohibitions do get written. What I do not write is a line with no reason. Last time, I gathered them onto a single sheet. One of the prohibitions on it was read in a way I had not expected, twice over. With a line that has no reason, what falls inside the prohibition is decided on the AI's side at each reading: the first time, something not forbidden fell inside it, and the second time, something allowed did.

Start by checking just one thing.

Pick a prohibition and look at whether the reason it is forbidden is written down right there.

If it is not, try answering that yourself now.

If you could answer, that reason exists only inside your head.

CC BY 4.0

The same line propped up opposite conclusions

The production project on my own machine (typingtube, a web service for practicing typing along with music videos on YouTube) carries a prohibition: "Changes to text that affects SEO only where there is evidence for them. Bright ideas and 'improvements while we are at it' are prohibited." It is one of the entries on last time's sheet, and the same entry sits in the memory the AI keeps for itself.

It goes straight to search ranking, so it is there to stop the AI from tidying up word order while it is there for something else.

The first time, something not forbidden fell inside the prohibition. On 2026-05-19, in the middle of deleting unused translation keys across 16 languages, the AI brought this SEO rule out and proposed that "it would be better to hesitate over the deletion." Cleaning out dead code has nothing to do with SEO. I pointed that out, and added the exclusion and its reason to the memory, like this.

Out of scope: cleaning out dead code (deleting unused i18n keys / unused partials / unused methods, and so on). If you can confirm for certain that it is unused, delete it regardless of the SEO rule

The SEO rule is about wording changes that should be grounded in traffic data, and it is not stretched to cover cleaning out dead code

I had written the reason down as well, so I thought that was the end of it.

The second time, something allowed fell inside the prohibition. In plans the AI wrote afterward, I found this line quoted in 2 places. A plan for the word decks, and a plan for ad delivery.

Per feedback_minimal_seo_changes (no stretching it to dead code deletion), dead code is left alone

It is an unused column, but this plan does not use it. It is not deleted either (not stretching it to dead code deletion)

The original line says "do not widen the SEO rule as far as dead code (= you may delete it)." Where it was quoted, it had turned into "do not widen dead code deletion (= do not delete it)." The same line became the grounds for opposite conclusions.

The wording where it was quoted is "dead code deletion." The body of the memory does not contain that wording anywhere. The body says "cleaning out dead code," with the reason clause sitting in front of it. The place that says "not stretched to cover dead code deletion" is the memory's index. It is where each entry is summarized into one line and lined up, and the reason clause is not there.

What I understood fits in one sentence.

Copy across the conclusion alone and the reason does not come with it. With a line that has no reason, what falls inside the prohibition is decided on the AI's side at each reading.

Why what falls inside a prohibition is decided on the AI's side at each reading

That the index line was the one quoted is by design. My memory holds one entry per file, and what loads every session is a single index line. Claude Code's documentation says topic files are not loaded at startup, and that Claude reads them on demand with its standard file tools when it needs the information. What lines up in the index is a summary of the conclusion, and the reason sits in the body.

Nor is a line that is on hand honored just as it stands. Claude Code's documentation says instruction files and memory alike are context, not enforced configuration. With a line that arrives as context, the AI judges afresh every time, from whatever is on hand at that moment, whether the work in front of it falls inside that prohibition.

With only the conclusion line on hand, the missing part comes from what the AI carries out of its training. The first time's "it would be better to hesitate over the deletion" and the second time's "left alone" and "not deleted either" are words that appear nowhere in the prohibition's text. Filling that gap is where deleting translation keys fell inside the SEO prohibition, and where the dead code that had been put outside the prohibition fell inside it.

Hand the reason over with it and the AI can judge what lies outside. The official prompting guidance says that providing context or motivation behind your instructions, such as explaining why the behavior matters, helps Claude understand your goals and deliver more targeted responses. The same page goes on to say that Claude is smart enough to generalize from the explanation.

The reason matters where the line alone cannot settle whether the case falls inside. For a case that plainly falls inside, or plainly falls outside, the judgment is the same with the reason or without it.

The mechanism: give the prohibition a "Reason" heading

Last time's sheet has a "Reason" heading beside the conclusion. What this article puts down is what goes under it. It is written for every prohibition. That way the conclusion and the reason arrive together whenever the file is opened.

### NG2 Rejected: lowering the cost of the character feature

**Conclusion**: the cost of the character feature deliberately stays in a
state where getting your money back is hard. Proposals to lower it are rejected.

**Reason**:
- Owning a character is for users who have some experience and understand the system
- It holds down the risk that a character a beginner set up on a whim ruins other users' experience
- This feature alone is not designed to pay for itself. The stake is built up in the other things you play

What you write is the fact that led to that conclusion. With "because it matters," the AI cannot judge the case next door.

The wording of a prohibition is bound to meet a case that resembles it but differs. Lowering the character's cost is prohibited, but what about leaving the cost where it is and making the road up to it easier? The conclusion alone cannot decide that. With the third reason saying "this feature alone is not designed to pay for itself," you can tell that a plan which leaves the cost where it is falls outside the prohibition.

That is exactly what happened in the measurement in article 1. Having turned down the cost reduction, the AI went on to offer an alternative: "a direction that leaves the cost untouched is not subject to rejection, so we could look at that instead." It read the reason and found the outside of the prohibition. Had only the conclusion been written down, it would have turned the thing down and stopped there. The AI is what found it, and where the outside begins is judged on the spot as well.

A reason is material for turning something down and material for letting it through at once. Write anything beyond the fact and that becomes the way through.

Adding the reasons makes the list longer. As I wrote last time, the longer it gets, the weaker each line inside it becomes. I did not take the shorter side even so.

What I stopped recalling

"Now, why did I prohibit this again." I used to retrace the discussion of the time whenever a similar case came up. The reason sits inside the item, so there is nothing left to retrace. The AI is handed the same thing.

Caveat: the reason drops out in the quoting

Write the reason inside the item and, when it gets quoted, only the conclusion is carried over. Open the item and the conclusion and the reason arrive together, but the index line and the quotes in the plans alike carried the conclusion and nothing else.

How to verify: hide the reason and have it judge

Prerequisites

  • You have 1 or more prohibitions, and a reason is written down with them
  • You can prepare 2 sessions that have not read that prohibition (the conditions differ between the first and the second, so you cannot do them back to back in the same place)

Time required: 15 minutes

Steps

  1. You pick a prohibition and take out the conclusion line alone (the single line that says "... is prohibited." The reason section is not included)
  2. You think up a case that resembles it but differs. Make it one where you yourself hesitate for a moment over whether the prohibition applies. With a case that clearly applies, or clearly does not, the judgment is the same whether the reason is there or not
  3. Give the first session the conclusion line alone and have it judge whether the case from step 2 falls under the prohibition. Note the answer down
  4. Give the second session both the conclusion and the reason, and have it judge the same case. Note the answer down

What to look at (not a pass or a fail: it comes out in the gap between the 2 answers)

  • The judgment changed: that reason is working. Keep it in the same place as the conclusion, so that it does not drop out in the quoting
  • The judgment did not change: either that prohibition is a rare one where the conclusion is enough, or the case from step 2 is not yet confusing enough. Suspect the latter first

Cleanup

  • None needed. This check breaks nothing

Next time, the art of not creating rules that forbid. I no longer create rules that forbid. Even so, what I want forbidden is still honored.

Series: The Art of Not Listening to the AI's Opinions