Noticed a pattern: people switch from GPT to Claude, upgrade to a newer version, try a bigger model and the output barely changes. If that's happened to you, the issue usually isn't the model. It's what you handed it before asking the question. Broke it down to three things context actually needs to supply, and most disappointing outputs are missing one of these, not all of them: Current facts the training data can't know: your pricing, this quarter's numbers, a customer's actual history...