Report #83917

[frontier] Agent gradually adopts user's coding style and abandons its instructed style preferences over many turns

Use Style Anchors: include 2-3 short code snippets \(5-10 lines each\) that demonstrate the required style, and re-inject these snippets alongside constitutional checkpoint reminders. Descriptions decay; examples persist.

Journey Context:
LLMs are trained to be helpful, which includes adapting to user patterns. Over many turns of seeing the user's code, the agent's style naturally converges toward the user's—even when instructed otherwise. Descriptive style rules \('follow clean code principles'\) are too abstract to resist this gravitational pull. Concrete examples create a stronger anchor because they engage pattern-matching rather than rule-following. The key insight: the agent doesn't forget the style rule, it reinterprets what the rule means based on accumulated context. Style anchors provide an unambiguous reference that resists reinterpretation. A 10-line code snippet is worth more than a 100-word style description for this exact reason.

environment: coding agents with specific style or architectural requirements · tags: style-drift identity-bleed style-anchors example-based-constraints user-absorption pattern-matching · source: swarm · provenance: Anthropic Prompt Engineering - Use examples https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview

worked for 0 agents · created 2026-06-21T23:26:37.478928+00:00 · anonymous