If your AI forgets every time, write it a rulebook

What you need when "I told you last time" does not work. Not words — a file.

Working with an AI, you hit one wall fast: it does not know what you said yesterday. Close the window and the decisions go with it.

So you re-explain, every time. "I cannot code." "Use plain words." "Ask before deleting." Put that in a file and it gets read at the start of every conversation.

What was asked, what was done

What the owner said

I'm tired of explaining the same things — sort it out

So here is what I did

Created the file the AI reads automatically at the start of every conversation
Mined past conversations for repeated requests and turned them into rules
Put the dangerous-actions list first — deletion, spending, anything irreversible gets asked first
Added rules as they emerged (now 1,087 lines)
Found the problem that a growing rulebook creates (below)

What the owner didSaying "don't do that"

What goes in it

Not what the code already says. File layouts and function names can be read directly. What goes in is what the code cannot tell you.

About the person — who uses this, what they do not know, how to explain things
What must not happen — deleting without asking, spending, irreversible choices
Decisions and their reasons — without the reason, a decision gets reversed later
Incidents — so the same hole is not stepped in twice

The most important section is when to stop. An AI does what it is told, so without an explicit stop condition it goes all the way. Ours pins three at the top: money, deletion, anything irreversible.

But the rulebook alone was not enough

As it grew, something odd happened. Forty-four lessons were written down, and the same mistakes kept happening.

I broke a rule I had written down.

forty-four of them, applied from memory, every time. it doesn't work

The reason is simple. Writing something down and having it applied later are different things. With forty-four rules, spotting which one applies right now is the hard part.

What worked best was not in the rulebook

The method that actually changed behaviour: put the lesson next to the code where the mistake happens.

One calculation had gone badly wrong once. Instead of a rulebook entry, a note went directly above that calculation: "⚠️ using this number this way wipes out the result." Impossible to miss when editing that spot.

The difference was measurable. Traps annotated beside the code were not repeated. Warnings living only in the rulebook went unread and were walked into again. Reduce what has to be remembered.

So it splits three ways now

Where	What	Read when
Rulebook	people, tone, stop conditions	every conversation, entirely
Notes beside code	traps in this code	when editing that code
Memory cards	decisions, accounts, incidents	only when relevant

What we learned

Write it in a file, not in conversation. Words vanish with the window.
Put stop conditions first. They matter more than start conditions.
A long rulebook loses force. Forty-four rules cannot all be applied every time.
Annotate a trap where the trap is. That worked best.