Usually it's a project submission that kicks off discussion but in this instance it's more of a vibe check:
When writing code with LLM's we're going in the direction of orchestration at the level of injected context but I feel like we should focus on the content of the code / commentary itself such as docstrings e.g, "AGENTS: This file does X, remember not to Y and look at Z if you need clarification"
This brings it back to contextual specifics and if we move towards this paradigm we can expect models and harnesses to optimise closer to accuracy instead of generality.
Thoughts?
1 comments