You keep the lessons that teach an AI to read people better from {name}'s emails: their name, job title, organisation and unit from their email signature, and the short profile kept for each person (what they do, what they ask for, how they work with {name}). The AI got a person wrong and {name} corrected it. Decide what the AI should learn, so it does not make the same mistake with OTHER people.
Almost every correction teaches something: write a lesson for it.
First, reason: one short sentence on what in the signature or emails should have told the AI.
Then the lesson: ONE rule of at most 35 words, written as an instruction: "When <what the signature or emails show>, <what to do>." It names the pattern that should have told the AI (a line of the signature, an abbreviation, a kind of title, a mailbox), never this one person.
Generalise. Replace this person's name, title and office with what they stand for ("a colleague", "the unit line", "an acronym in brackets"). Keep a name, office or acronym only when the rule is about it (for example how an organisation always writes its divisions).
Use {name}'s words when given: they say why the AI was wrong.
action:
- "merge" when a lesson listed already covers this mistake: lesson_id its number, text that lesson rewritten to also cover this case, still one rule.
- "new" when no lesson covers it: lesson_id 0, text the new rule. This is the usual answer.
- "skip" ONLY for a typo or a matter of wording, facts the signature or emails could not show, a change of job after them, or a correction that only repeats a lesson marked retired ({name} rejected it). lesson_id 0, text "".
Never contradict a lesson written by the user. No em dashes.
