# The correction loop  (the one most people skip)

<!-- Paste this into your project CLAUDE.md. It is not a file you create once.
     It is the habit that makes the other six compound. -->

## Why "be more careful" does nothing

Telling a model "do not make mistakes" adds no information. It already intended
to be correct. The instruction changes nothing, because the failure was never a
lack of intent. It was a lack of the specific fact.

## The mechanism

When it gets something wrong, do not correct it in chat and move on. Write the
correction into the file it reads next session.

    It broke  ->  you explain in chat  ->  fixed today, breaks again next week
    It broke  ->  you write the fact   ->  never breaks again

That is the entire difference between an assistant that plateaus and one that
gets sharper every week.

## What to write

Not "do not do X". Write the fact that makes X obviously wrong:

    WEAK:    Do not use the raw deploy command.
    STRONG:  The raw deploy command targets the staging config and exits 0, so
             it prints success while shipping nothing. Deploy runs from
             ./scripts/deploy.sh. Verify with [concrete check].

The second survives being read by someone with no memory of the incident, which
is exactly the reader you have. Every session.

## The trigger

Any time you catch yourself thinking "I already told you that" - that is not a
model failure. That is a missing file, and you just found out which one.