Research on the Courts of Appeals has been limited by the lack of a universe-level database covering the courts' 440,000 published opinions from 1892 to 2025. Using a multiple-step large language model approach, we create such a database, including metadata (e.g., the judges involved in each case), summary information (e.g., substantive and procedural issues raised, litigant types), and the decision's ideological direction (liberal or conservative). We validate our database against multiple sources, showing that our LLM output matches human coders 85-90\% on key summary variables and 80\% of the time on ideological direction. The new database will enable much more comprehensive analyses of the Courts of Appeals over time; as examples, we extend existing work on panel effects and ideological decision-making. More generally, the approach we take provides a pipeline for converting expert-written codebooks into machine-extractable facts, which has relevance for computational social science beyond judicial politics.