Stop Repeating Yourself to Your AI

“Use British English. Follow our conventions. Don’t touch the migrations. Check the design system first.” If you type some version of that every session, the problem is not your memory — it is that the instructions live in the conversation instead of the project.

What the repetition actually costs

It looks like a small tax — thirty seconds at the start of a session. The real cost is what happens on the sessions where you forget.

The preamble tax

Every session opens with the same paragraph, and every session it is slightly different from the last one.

The forgotten session

The one time you skip it, the agent does the thing you always tell it not to — and you find out in review.

The team divergence

Everyone writes their own preamble, so the same codebase gets built to five slightly different sets of conventions.

Move it out of the conversation

1. Write down what you always say

Open your last few sessions and look at the preamble. Whatever you typed more than twice is a rule. Most people find their first three in about a minute.

2. Say how strongly you mean each one

"Never commit credentials" is a prohibition. "Prefer date-fns" is a preference with a legitimate exception. "Check the design system first" is a pointer at a document. The type is what stops your agent either ignoring rules or stopping to ask about everything.

3. Attach it and stop typing it

Attach the ruleset to your projects. From then on every session — yours and your teammates’ — starts with those instructions already in place, and you can see exactly which ones were in effect.

You do not have to write them by hand. Tell your assistant “remember that we always use British spelling” and it will create the rule, choose the right type and record why. Every rule an AI writes is versioned and logged, so you can see what changed and undo it.

Common questions

Why does my AI forget my preferences between sessions?

A chat session has no memory of the last one. Anything typed as a message is gone when the session ends, so preferences stated in conversation must be restated every time. Store them outside the conversation and they stop disappearing.

Is putting instructions in a file enough?

For one person on one repository, often yes. It breaks down across several repositories or several people: copies drift, nothing can be made non-negotiable everywhere, and you cannot tell afterwards whether a rule failed or was never delivered.

What belongs in standing instructions, and what stays in chat?

Standing instructions are for what is true every time — conventions, prohibitions, house style, pointers to documents. Today’s specifics stay in the conversation. Read it aloud: if it would still be true next month on a different task, make it a rule.

Will this fill up my context window?

Only if you let it sprawl, which is why there are limits — twenty rules in the baseline, fifty per ruleset. Keep each to a sentence, and for anything longer use a consult rule pointing at a document that loads only when relevant.

Write the preamble once

Then never type it again.

Start free