After the new system failed, a legacy project maintenance requirement brought AI down to earth

The third installment of "Embedded AI Implementation Notes" records that after the new system failed, the author spent about a week building a legacy project maintenance skill library with only five skills and around seven hundred lines of code, and ran it successfully in a real change. The process sets up two manual gates: twenty-one assumptions confirmed one by one, and the modification plan must be explicitly approved; the first replay verification generated forty-five test cases, with zero compilation errors and zero new issues from static checks. The author suggests starting with maintenance-type work and placing manual confirmation at decision points such as assumptions and plans.