What rules does my AI need – and do I have to know them myself?

Facts as of October 17, 2026

Few, and each worded so that a violation can be detected: “Be careful” can't be broken, “no transfer without a checksum shown first” can. The best rules come from an incident. And yes, you have to know them: a rule that only exists in a file works exactly as far as the AI follows it.

In detail

Word rules so they can be checked

A rule like “be careful” can't be broken, because you can never decide whether it was broken. A checkable rule says what has to be in the answer. Examples from the rulebook of Thomas J. Weiss:

Every factual claim that supports a recommendation carries a source or the label “assumption.”
No statement with a time reference (“today,” “still,” “by now”) unless the time was checked in the same answer.
A proposal that touches tools only counts with a test that was actually run, including visible output.
If the AI notices ongoing damage, it stops it without asking; starting and deleting still require approval.

Learn from incidents

Record every violation: date, rule, what happened, damage, lesson. Not as an apology, but as evidence. Many useful rules have exactly this kind of origin, and later that origin is the reason not to delete them.

Rules alone don't protect you

A rule can be in the file and still not be there at the decisive moment, see How do I stop the AI from deleting something it shouldn't? More text doesn't help then. What helps is a checkable stop before the action – a list, an OK, a cross-check.

Clean up

The file that every session reads at the start contains only rules that apply; whatever is history or done goes to the archive. Before a rule is changed or deleted, the AI checks whether it is still current and what purpose it serves, and presents both.

Know them yourself

If you only work in the chat, you don't read the rulebook. Then it exists for the AI, not for you: you can check neither whether it follows the rules nor whether they are even right. So every now and then, have the AI present the rules that remain in force, each with its origin and reason. That way you can delete instead of having to remember.

Knowing also means reading as much as possible of what the AI gives back to you. If you can't manage all of it, at least read the thread carefully and ask as soon as something there contradicts itself. Only a look into the store, on the other hand, shows whether a result has really been stored, see How can I tell whether the AI has really done, saved or checked something?

In practice

September 20, 2026 shows how a rule grows out of an incident: his Claude instance had read the clock at 1:35 a.m. and, one round later, recommended starting a project “not tonight.” In between, he had slept; it was 9:19 a.m. Since then the rule is: time references only when measured.

When he looked back on a play he had written with an AI for his parish, it turned out that he had never read the project's rulebooks:

I can't say exactly in hindsight, since I didn't read the rulebooks either, but mostly just used the chat.

(Original: „Ich kann das im Nachhinein nicht mehr genau sagen, da ich die Regelwerke auch nicht gelesen habe, sondern meistens einfach nur den Chat benutzt.“)

One of the AI's roles was described incorrectly in the rulebook for three weeks, and nobody noticed. His answer to the question of what should become of the rules:

I'd say we don't need the discarded rules anymore, but above all what's left over. Of the ones left over, though, definitely their history and the reason why they exist.

(Original: „Ich würde sagen, die verworfenen Regeln brauchen wir nicht mehr, sondern vor allem das, was übrig geblieben ist. Von den übrig gebliebenen aber durchaus ihre Entstehungsgeschichte und der Grund, warum es sie gibt.“)

In October 2026 he answered the question of whether you have to know your AI's rules yourself once more. He doesn't read what is in the knowledge store, but he reads what is in the thread very carefully; if there is a contradiction there, he asks:

The answer is still yes, you have to know the AI's rules yourself and, if possible, also read everything the AI gives back to you. I don't read what's in the vault, though, but I read what's in the thread very carefully. I can't manage more. If there's a contradiction in it, then I ask again.

(Original: „Die Antwort ist immer noch ja, man muss die Regeln der KI selbst kennen und möglichst auch alles lesen, was die KI einem zurückgibt. Ich lese allerdings nicht, was im Vault steht, sondern ich lese das, was im Thread drin steht, sehr genau. Mehr schaffe ich nicht. Wenn ein Widerspruch drin liegt, dann frage ich nochmal nach.“)

Sources

Note „Sieben Mechaniken“, replicapong.de, 11.10.2026
Note „Zusammenarbeit an Texten“, replicapong.de, 11.10.2026
Note „Vier Grundsätze und sechs Schichten“, replicapong.de, 11.10.2026
Interview on the review of a theater project, 02.09.2026
Review of the rules of the same project, 02.09.2026
Dictation by Thomas J. Weiss comparing with the theater project, 17.10.2026

Cite: Thomas J. Weiss: “What rules does my AI need – and do I have to know them myself?” replicapong.de, as of October 17, 2026. https://replicapong.de/en/question/rules-for-your-ai/ – CC BY 4.0