Definitely, and hence the reason that structuring requests/responses and providing examples for smaller atomic units of work seem to have quite a significant effect on the accuracy of the output (not factuality, but more accurate to the patterns that were emphasized in the preceding prompt).

I just wish we could more efficiently ”prime” a pre-defined latent context window instead of hoping for cache hits.