Reinforcement Anchors Beat Emphasis: Compressing a Production System Prompt

Production system prompts bloat by a predictable mechanism. The model does something wrong, so you add an instruction telling it not to. When that doesn't stick, you add a more forcefully worded one. NEVER do X. In capitals. With exclamation points. Salesforce found this same escalation across 20,000 enterprise agent deployments, and found it does not work: an LLM does not process typographic emphasis the way a human reader does. Capitalization and punctuation are just more tokens, not a signal that reliably overrides competing considerations during generation. So the instruction fails again, another one gets appended, and the prompt accumulates. I took one production prompt from roughly 4,000 words to roughly 1,300 and it got more reliable, not less. That is only surprising if you believed the length was buying compliance in the first place.

What actually carries a constraint is position, not volume. Attention has a measurable front-and-back bias: RoPE, the positional encoding most current models use, decays in a way that puts tokens far from both ends of the sequence into a systematically lower-attention zone, and retrieval accuracy for a fact placed mid-context drops by more than 20 points compared to the same fact at the start or end. A constraint's location is load-bearing in a way its wording is not. So the rewrite was tiered rather than shortened: identity and non-negotiable constraints at the edges, task detail in the middle, and reinforcement anchors placed to survive attention decay across a long multi-turn conversation rather than only the first exchange. The other half of the compression was subtraction. Salesforce's corollary is that anything you can draw as a flowchart belongs in code, not in a prompt, because code executes identically every time and no wording does. A context window is an attention budget for the run, not a junk drawer for everything that once went wrong.