[Bug] Model ignores Chilean Spanish tuteo instruction, drifts to Rioplatense voseo in long contexts

Open 💬 0 comments Opened Jun 10, 2026 by danionpar

Bug Description
Title: Model drifts from instructed Chilean Spanish (tuteo) into Rioplatense voseo mid-session

Environment:

  • Model: Opus 4.8 (claude-opus-4-8[1m])
  • Surface: Claude Code CLI
  • System prompt explicitly instructs: "In Spanish, use neutral Chilean Spanish with tuteo (tú). No voseo, no Rioplatense." — present in both ~/.claude/CLAUDE.md and the active output style.

Expected: All Spanish replies stay in neutral Chilean tuteo (tú / tienes / quieres) for the entire session, as instructed.

Actual: Replies start correctly in Chilean tuteo, then drift into Rioplatense voseo (vos / tenés / querés / dale / mirá) partway through the conversation. The drift correlates with longer contexts — the further the conversation
progresses, the more the explicit system-prompt instruction is ignored.

Impact: The user is Chilean. Being addressed in Argentine voseo despite an explicit, repeated instruction is jarring and reads as the model overriding a stated user preference. A negative instruction at the top of context is not holding
against the model's base tendency toward over-represented Rioplatense Spanish.

Repro:

  1. Set a system prompt / output style instructing neutral Chilean Spanish with tuteo, explicitly forbidding voseo/Rioplatense.
  2. Hold a multi-turn Spanish conversation of moderate length.
  3. Observe register shift from "tú/tienes" to "vos/tenés" without any user prompt change.

Suggested fix: Strengthen adherence to explicit regional-Spanish register instructions, particularly resistance to register drift over long contexts. The model should not fall back to a default Spanish variant when a specific one is
instructed.

Environment Info

  • Platform: darwin
  • Terminal: ghostty
  • Version: 2.1.170
  • Feedback ID: 827e40ed-930e-4a79-8cba-e2905726f9ca

Errors

[]

View original on GitHub ↗