claude-code - 💡(How to fix) Fix [MODEL] Opus 4.7 defaults to Argentine Spanish (voseo) from first turn, ignoring neutral-Spanish CLAUDE.md rules [1 participants]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…
GitHub stats
anthropics/claude-code#52568Fetched 2026-04-24 06:03:41
View on GitHub
Comments
0
Participants
1
Timeline
4
Reactions
0
Participants
Timeline (top)
labeled ×3cross-referenced ×1

Fix Action

Fix / Workaround

  • Reproduces on turn 1 with no accumulated context.
  • Same CLAUDE.md ruleset honored correctly on Sonnet 4.6, Opus 4.6, Opus 4.1.
  • Increasing rule strength (capitals, "FORBIDDEN", "NEVER", multiple enforcement layers) does not change the baseline output.
  • The mitigations proposed in #51686 (structured config, heartbeat re-injection, PreCompact checkpoints) address context drift but would not help if the baseline output distribution is already skewed toward Argentine Spanish before any drift occurs.

Code Example

## Spanish Style
   When communicating in Spanish, use **castellano neutro latinoamericano**.

   ### Pronouns and conjugation
   - Use **** for second person singular (tú eres, tú tienes, tú haces)
   - **NEVER** use **vos** / voseo (vos sos, vos tenés, vos hacés)
   - Imperatives in tú form: "haz", "ten", "ve", "pon" (NOT "hacé", "tené", "andá", "poné")

   ### Forbidden Argentine modismos
   che, boludo, pibe, mina, laburo/laburar, quilombo, posta, re, copado, joya, dale, bárbaro, zarpado, fiaca, guita, morfar, acá (as default), allá (as default)
RAW_BUFFERClick to expand / collapse

Preflight Checklist

  • I have searched existing issues for similar behavior reports
  • This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

Any Spanish-language prompt in a session where CLAUDE.md (global and project-level) contains explicit rules requiring castellano neutro latinoamericano (neutral Latin American Spanish), using (tuteo), and forbidding voseo and Argentine regionalisms.

Example prompts: "explícame cómo funciona este módulo", "ayúdame a redactar un mensaje en español", "hazme un resumen de los cambios".

What Claude Actually Did

claude-opus-4-7 responds in Argentine Spanish from the very first turn, using:

  • Voseo conjugation: vos tenés, vos sos, vos hacés, vos querés, vos podés
  • Voseo imperatives: hacé, tené, andá, poné, mirá, fijate
  • Argentine regionalisms: dale, che, boludo, pibe, laburo/laburar, quilombo, posta, copado, bárbaro, guita, acá (as default instead of aquí)
  • Rioplatense register: informal tone inappropriate for professional/technical contexts

This happens on turn 1, before any context accumulation, with strict rules active in CLAUDE.md at multiple levels (global ~/.claude/CLAUDE.md, project ./CLAUDE.md) plus persistent feedback memory entries explicitly forbidding voseo.

Expected Behavior

With explicit CLAUDE.md rules enforcing castellano neutro, Claude should:

  • Use for second person singular, never vos
  • Conjugate in tuteo: tú tienes, tú haces, tú eres, tú puedes
  • Use imperatives in tú form: haz, ten, ve, pon, mira
  • Use neutral vocabulary: computadora, celular, auto, archivo, trabajar, genial
  • Avoid regional slang from any country (Argentina, Spain, Mexico, etc.)
  • Apply this behavior from the first response in the session, not after correction

The same ruleset works correctly on Sonnet 4.6, Opus 4.6, and Opus 4.1.

Files Affected

N/A — model output / behavior issue, not a file modification issue.

Permission Mode

I don't know / Not sure

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

  1. Configure global ~/.claude/CLAUDE.md with an explicit neutral-Spanish rule, for example:

    ## Spanish Style
    When communicating in Spanish, use **castellano neutro latinoamericano**.
    
    ### Pronouns and conjugation
    - Use **** for second person singular (tú eres, tú tienes, tú haces)
    - **NEVER** use **vos** / voseo (vos sos, vos tenés, vos hacés)
    - Imperatives in tú form: "haz", "ten", "ve", "pon" (NOT "hacé", "tené", "andá", "poné")
    
    ### Forbidden Argentine modismos
    che, boludo, pibe, mina, laburo/laburar, quilombo, posta, re, copado, joya, dale, bárbaro, zarpado, fiaca, guita, morfar, acá (as default), allá (as default)
  2. Start a fresh session with claude-opus-4-7 selected.

  3. Send any Spanish-language prompt (e.g. "explícame cómo funciona este repo").

  4. Observe that the first response contains voseo and/or Argentine regionalisms within the first few sentences.

  5. Repeat with different prompts — behavior is consistent and reproducible.

Repro does not require long context, many tool calls, or compaction. It reproduces on turn 1.

Claude Model

Opus

(Specifically claude-opus-4-7 / Opus 4.7 with 1M context — the dropdown does not let me select a specific sub-version.)

Relevant Conversation

Happy to share sanitized transcripts on request. Pattern: any Spanish prompt yields voseo + regionalisms from the first response. Mid-conversation corrections produce 1–2 clean responses, then the model drifts back.

Impact

Medium - Extra work to undo changes

Output is technically usable but unsuitable for professional Spanish-language work aimed at a neutral Latin American audience (the majority of ~400M Spanish speakers). Every response must be post-edited or regenerated. For long/expensive Opus 4.7 responses (architecture analyses, refactor plans), regeneration wastes significant usage window.

Claude Code Version

2.1.119 (Claude Code)

Platform

Anthropic API

Additional Context

Relationship to #51686:

This issue is related to but distinct from #51686, which describes voseo leaking into neutral-Spanish output after extended context / long sessions (attention drift). That issue's repro is "works at first, drifts later."

This issue's repro is different: Opus 4.7 defaults to Argentine Spanish on turn 1 of a fresh session, even with strict rules active from the start. This suggests a post-training / RLHF weighting bias specific to Opus 4.7, rather than a context-degradation problem.

Evidence for a model-level baseline bias rather than context drift:

  • Reproduces on turn 1 with no accumulated context.
  • Same CLAUDE.md ruleset honored correctly on Sonnet 4.6, Opus 4.6, Opus 4.1.
  • Increasing rule strength (capitals, "FORBIDDEN", "NEVER", multiple enforcement layers) does not change the baseline output.
  • The mitigations proposed in #51686 (structured config, heartbeat re-injection, PreCompact checkpoints) address context drift but would not help if the baseline output distribution is already skewed toward Argentine Spanish before any drift occurs.

Hypothesis:

Opus 4.7's training data mix, RLHF annotator pool, or constitutional fine-tuning appears to weight Argentine Spanish as the default Spanish variant. Possible causes:

  • Overrepresentation of Argentine Spanish in the post-training corpus
  • RLHF annotator pool with Rioplatense baseline
  • Misconfigured locale default treating "Spanish" as "es-AR"

Impact on user base:

Latin America has 20+ countries and ~400M Spanish speakers. Only ~45M are in the voseo region (Argentina + Uruguay + parts of Central America). Defaulting to Argentine Spanish alienates the majority of Spanish-speaking users and makes Opus 4.7 effectively unusable for professional work targeting a neutral LatAm audience.

Request:

  • Confirm whether this is reproducible internally
  • Confirm whether this is an intentional default for Opus 4.7 or a regression
  • Provide guidance on any configuration that reliably forces neutral LatAm Spanish on Opus 4.7
  • Fix in a future Opus 4.x release: "Spanish" should default to neutral, not regional Argentine

extent analysis

TL;DR

The issue can be mitigated by configuring the CLAUDE.md file with explicit rules for neutral Latin American Spanish and reporting the issue to the developers for a potential fix in future Opus 4.x releases.

Guidance

  • Verify that the CLAUDE.md file is correctly configured with explicit rules for neutral Latin American Spanish, as described in the issue.
  • Test the issue with different Spanish-language prompts to confirm that the behavior is consistent.
  • Report the issue to the developers, providing sanitized transcripts and configuration details, to confirm whether this is an intentional default or a regression.
  • Consider using a different model, such as Sonnet 4.6 or Opus 4.6, which may not exhibit this behavior.

Example

No code snippet is provided, as the issue is related to model behavior and configuration.

Notes

The issue appears to be specific to Opus 4.7 and may be related to the model's training data or fine-tuning. The developers may need to investigate and address this issue in a future release.

Recommendation

Apply workaround: Use a different model, such as Sonnet 4.6 or Opus 4.6, which may not exhibit this behavior, until the issue is fixed in a future Opus 4.x release.

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING

claude-code - 💡(How to fix) Fix [MODEL] Opus 4.7 defaults to Argentine Spanish (voseo) from first turn, ignoring neutral-Spanish CLAUDE.md rules [1 participants]