The gist: - At scale prompt caching is your main money saver. - Prompt caching and dynamic tool discovery bite each other. Tools tend to be defined in the system prompt. So if you use dynamic tool discovery you invalidate the cache, which is expensive at scale. If you have control over the model, you can mask (most of us can't so @AnthropicAI, please build this :D) - Use the file system as context. I would rather say delegate work to other things. Once you want to use the context, you still need to read it in. And to write the context, you actually need to have the content in the context window. If you delegate to subagents, you can actually keep the context window clean. Also take advantage of the tendencies of the model. Cursor likes to read in 200 line chunks, Amp is more dynamic (10 / 50). Keep your file sizes in line with these sizes.