Back to all notes

Design note / Active prototype

Practice, not self-criticism

Why a system that notices its own habits should not be told about them.

Anything keeping notes on its own behavior will end up writing down mostly failures. A reply that lands leaves no trace. A correction leaves an obvious one. Build the record from that and you get a list of faults and nothing else.

Read it back before every reply and it stops being a description. Tell a language model it tends to over-explain, put that in front of it each time it speaks, and you have not fixed the habit. You have taught it.

Answer the habit, do not describe it.

So the notes are kept, and what gets read back is different. Something that works can safely be described, because describing it encourages more of it. Something that is not working gets one specific thing to do instead, and the description of the failure stays out of the way.

A complaint with no answer attached comes through as nothing at all, which is the right outcome. An unanswered fault, read aloud every time, is worse than silence.

See whether it helped.

A change is only worth anything if something moves. So it compares how often the problem shows up now against how often it showed up before, in proportion rather than in raw counts, because a busier week produces more of everything without anything having gotten worse.

Something that helps twice becomes permanent and drops out of the way. Something that changes nothing gets withdrawn and reworded. A second miss is written down as unresolved, which is more honest than calling it impossible, because a number that did not move cannot tell a badly worded fix from a measurement too rough to see one.

Leave the deleting to a person.

Adopting, retiring, and rewording all happen quietly in the background. Deleting does not, because it is the one step nothing later can undo. That waits for you.

And sometimes it is not a habit at all. When the evidence points at something broken rather than at the way it talks, it is presented as a bug and kept well away from the self-record. Noticing has turned out to be a decent way to find real problems.

Explore the prototype