[BUG] Claude Code ignores CLAUDE.md language and formatting instructions when generating documents

Resolved 💬 4 comments Opened Feb 16, 2026 by msouga Closed Mar 20, 2026

Preflight Checklist

  • [x] I have searched existing issues and this hasn't been reported yet
  • [x] This is a single bug report (please file separate reports for different bugs)
  • [x] I am using the latest version of Claude Code

What's Wrong?

  1. I asked Claude Code to generate three technical manuals (~250 lines each) as Markdown files.
  2. Claude generated all three documents in Spanish but without any diacritical marks: no tildes (wrote "accion" instead of "acción"), no eñes where needed, no opening punctuation marks (¿, ¡).
  3. I had to explicitly point out the issue and request a full rewrite of all three files.
  4. Claude rewrote all three documents correctly on the second pass.

What Should Happen?

When CLAUDE.md specifies language and formatting rules (diacritical marks, specific punctuation, etc.), Claude Code should apply them from the very first generation, especially for large content blocks like full documents.

Error Messages/Logs

Steps to Reproduce

  • Claude Code CLI
  • Model: claude-opus-4-6
  • OS: macOS (Darwin 25.3.0)

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.42

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

This is not a one-off issue. The model seems to default to "ASCII-safe" Spanish (no diacritics) when generating large volumes of text, even when explicitly instructed otherwise. The problem is particularly costly because documentation tends to be long, so a full rewrite means a significant amount of wasted tokens.

A possible mitigation would be for the model to do a quick self-check against CLAUDE.md language/formatting rules before outputting large documents, rather than relying on the user to catch and report the issue after the fact.

  • ~750 lines of documentation written twice (once incorrectly, once correctly).
  • Roughly 50% of session credits wasted on content that had to be discarded and regenerated.
  • User frustration: the instructions were clear and already present in CLAUDE.md. The user should not have to remind the model to follow its own project instructions.

View original on GitHub ↗

This issue has 4 comments on GitHub. Read the full discussion on GitHub ↗