How to Stop AI Forgetting Your Rules

Rules have a half-life

You set the rules at the start of the session, carefully, in one tidy block. Then you did an hour of real work, and the rules did what rules do in a long conversation: they faded. Not all at once, and not because anything went wrong. They simply moved further and further from the sentence being written now.

So stop thinking of your constraints as settings you configured once. Think of them as a signal that weakens with distance, and needs a repeater.

Fix one: restate at decision points

You do not need to repeat everything constantly. You need to repeat the relevant rule at the moment it is about to be tested.

A decision point is any message where the assistant is about to produce something that a colleague might actually see: a draft, a summary of findings, a recommendation, a rewrite, an email. Those are the moments where a forgotten constraint becomes a real mistake rather than a private one.

The move takes one line. Before the request, restate only the rules that bear on this specific output.

Before you draft this, the rules that apply here:

Findings only. No recommendations, no next steps, no suggestions.
Every claim needs a page reference from the material I supplied.
If the source does not settle something, write UNCLEAR. Do not estimate.
Now draft: [your request]

Three lines, and they sit immediately next to the work rather than an hour behind it.

Fix two: make it repeat the constraints back

Before anything large or expensive, force the assistant to reconstruct the rules in its own message. This does two things at once. You find out whether the rules survived, and if they did, they are now freshly restated in the most recent part of the conversation, which is exactly where they carry most weight.

Do not start the task yet.
First, list every rule and constraint currently in force for this piece of
work, numbered, in my wording. Include anything I set earlier that still
applies. Then state in one line what you understand the task to be.
I will confirm or correct before you begin.

Read what comes back properly. Two things are worth noticing. Missing rules tell you what has already decayed. Invented rules are just as important: an assistant that lists a constraint you never set has started building on its own assumptions, and everything downstream inherits them.

Fix three: keep a standing brief outside the chat

Write your rules once, in a note on your own machine, and paste the block in again whenever a session gets long or whenever you start a new one. Whether you work in Gemini, ChatGPT or Claude, the brief belongs outside the chat window, because that is the only copy that does not decay. Keep it short enough that pasting it never feels like a chore, because a brief you resent pasting is a brief you stop pasting.

STANDING BRIEF. These rules apply to everything in this conversation and
override anything we drifted into earlier.
Role: [what you are acting as, one line]
Audience: [who reads the output]
Always:

[rule]
[rule]
Never:
[rule]
[rule]
Uncertainty: if the material I supplied does not settle a point, write
UNCLEAR and stop. Do not estimate, and do not fill the gap from general
knowledge.
Confirm you have these, then continue with: [current task]

The phrase "override anything we drifted into earlier" is doing real work. Without it you have added rules to a pile that already contains contradictions. With it, you have said which version wins.

Fix four: start again, on purpose

Sometimes the thread is poisoned. A wrong figure got into an early answer and now reappears in every draft. A misunderstanding about scope has been quietly baked into six outputs. You correct it, it agrees warmly, and three messages later it is back, because the mistake is still sitting in the conversation being read as history.

Fighting that is a bad trade. You will spend more effort arguing than restarting, and you will never fully trust the result.

Starting fresh is not failure and it is not lost work. Take the standing brief, take the source material, take a clean summary of the decisions you have made, and open a new conversation. The trick is to carry the conclusions across without carrying the contamination.