attac-t Skill: Retrospect
"A fault a tool caught is a fault you did not."
Being corrected is not the problem. Needing to be corrected twice for the same reason is.
The four
| What did I claim that was not true? | say the claim, not the topic |
| Who caught it? | a gate, a test, a person. A person is the expensive one |
| What would have caught it earlier? | a check, a question, a habit. Name the cheapest |
| Is it the same as last time? | a repeat is a missing mechanism, never a lapse |
The fourth is the whole point
One wrong call is a mistake. The same wrong call twice is a rule nobody wrote.
Read one sample and call it a state. Guess a cause instead of measuring. Say a thing is proved when a test skipped. Each of those happened more than once in a day, and each was cheap to catch and was not caught — because nothing looked back.
What comes out
One of three, and say which:
| a fix | the fault is gone and something proves it |
| a check | a person cannot be the thing that notices this |
| a note | it was a one-off, and here is why you believe that |
A note is the weakest and the most common, so use it least. If you cannot say why it will not recur, it is a check you have not written.
Where it lands
Wherever the work lives. A run's own record when Floor is installed. An observation when it is
worth remembering — craft-observation owns that. A commit message otherwise.
Not a session note. A learning that dies with the harness was never a learning.
When to run it
When a piece of work ends. When a check goes red. When a person tells you something you should have seen.
That last one is not optional. It is the signal that something is missing, and the only one that arrives for free.