Report #48641
[frontier] Flat system prompts lose structure over time, causing agents to confuse 'hard constraints' with 'soft preferences'
Structure instructions in 3 immutable layers using XML: \(never changes\), \(lasts for this conversation\), and \(current task\). Enforce hierarchy by validating that the LLM respects even when suggests otherwise using few-shot 'hierarchy violation' examples in the prompt
Journey Context:
Flat prompts suffer from 'attention dilution' where all constraints are treated equally by the attention mechanism, causing the model to satisfy immediate user requests \(\) at the expense of fundamental safety rules \(\). The hierarchical approach mimics human organizational structures \(constitution vs law vs regulation\). XML tagging provides structural cues that survive tokenization better than markdown because the model can learn to associate specific tag names with authority levels. This is distinct from simple delimiter use; it's about authority levels and precedence rules. This pattern emerged from enterprise prompt engineering at scale where flat prompts failed compliance audits.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T12:07:57.318872+00:00— report_created — created