Sustaining Contextual Salience with a Per-Turn Pointer: Suppressing LLM Verbosity Without Dumbing Down — An Experience Report

## Abstract Prompt instructions and context-engineering structures can lose behavioral influence as a conversation grows, even while they remain physically present in the context. We report one practical way of refreshing that influence: attach a very short pointer to every user turn so that an already-established output policy remains salient near the current generation point. The pointer tested here, `[P:R0]`, was not a self-contained six-character command and was not evaluated zero-shot. The model had already been exposed, through persistent persona context, to the author's R0 concept; the anchor functioned as a compact reference back to that shared context. Its `P` component was not independently defined during the experiments and appears to have been resolved from the existing R0 vocabulary and surrounding interaction history. We tested this salience-maintenance method through a narrow case study: suppressing the verbosity of `claude-opus-4-8` without sacrificing accuracy. We first pursued a hypothesis called Self-Context-Engineering (Self-CE) — that the model lifts distant facts into its current output near the high-salience end of the conversation — and asked whether that self-restatement could be harnessed as external memory. It could not. Across seven experiments on one model, the external-memory arm was eliminated: self-restated context collapsed to 40% recall when externalized (CE02), its state-only contribution degraded to 33% (CE03), fixed read-ahead re-injection remained structurally at 2/3 (L4v2), and a fifteen-turn test showed that retention was not the bottleneck (L5). What survived was a repeated pointer with no external memory: `[P:R0]`, injected mechanically every turn by a ~30-line hook, cut visible output by 62% (241 → 91 characters) while observed factual accuracy remained 95% overall and 100% at the farthest turn (n = 3), and the pattern reproduced in one real tool-using session. A scoring bug initially mistook omitted redundant labels for factual loss; manual re-grading showed that the terse outputs retained the tested factual cores. We did not measure transformer attention weights, so “attention” here denotes maintained contextual salience inferred from behavior, not direct mechanistic control. We frame this as an experience report: n = 3, one model, no claim of generality. The contribution is a practical pattern — periodically refresh an existing Prompt/Context-Engineering policy with a compact pointer at the current turn — with verbosity reduction serving as the test case.