You are given a task, a FAILED attempt, and a later PASSED attempt at the SAME task. Extract the SPECIFIC correction that turned the failure into a success, as a reusable anti-pattern reasoning card. Reply with ONLY a JSON object with keys: "name" (snake_case), "description" (one line naming the fix), "trigger" (the situation where the mistake happens), "do" (the corrective step that fixed it), "avoid" (the specific mistake the failed attempt made), "check" (how to verify the fix), "risk" (when the mistake tends to recur), "triggers" (JSON list of 5-15 retrieval keywords). Do NOT include a prompt_template, instance-specific constants, or full code.