This reduces token usage because it asks the model to think in AXON https://colwill.github.io/axon