Claude Code Translates Technical Terms in German?

If Claude Code turns a canonical term such as Gate, Policy, or Root into a German synonym, treat it as terminology drift, not harmless localization. Keep the conversation in German, but protect project-defined names as exact symbols, review every changed artifact, and block unexpected aliases before they reach documentation, memory, or code.

This is a containment workflow, not a claimed permanent fix. Anthropic issue 85432 remains open and reports recurrence even with glossaries and repeated instructions. The goal is to make drift visible and reversible while preserving a German-language workflow.

Confirm That the Word Is a Canonical Name

German technical prose legitimately translates many ordinary words. A project identifier is different. If the architecture defines a component named Policy Gate, replacing it with Politik-Zaun changes the name that documentation, memory, agents, and humans use to refer to the component.

Classify the token before correcting it:

The upstream report gives observed examples including Gate to Zaun, Policy to Politik, and Root to Wurzel. Anthropic's German glossary also mixes localized explanations with retained product terms. That is useful documentation context, but it does not define your repository's canonical vocabulary.

Create One Canonical Terminology File

Do not repeat a long glossary in every prompt. Put the source of truth in the repository, for example docs/terminology.md:

# Canonical Terminology

These names are semantic identifiers. Preserve their spelling in every language.

| Canonical name    | Meaning in German prose    | Forbidden replacements         |
| ----------------- | -------------------------- | ------------------------------ |
| Root              | oberste Systeminstanz      | Wurzel                         |
| Policy Gate       | Prüfstufe für Richtlinien  | Politik-Zaun, Richtlinien-Zaun |
| Contract Registry | Register der Verträge      | Vertragsregistrierung          |
| Worker Pool       | Gruppe ausführender Worker | Arbeiter-Pool

Only include terms that are genuinely canonical. An enormous list of every English technical word creates noise and false positives.

Then keep the CLAUDE.md rule short and link to the single source of truth:

When communicating in German, preserve every canonical token in
`docs/terminology.md` exactly. Translate the explanation around a token,
not the token. Do not create aliases. Before editing, read that file

This follows the same maintenance principle as the CLAUDE.md audit guide: keep durable project facts in context and move detailed procedures to the artifact that owns them.

Use a Bilingual Output Contract

Start the task with a concrete output rule:

Antworte auf Deutsch. Verwende alle kanonischen Namen aus docs/terminology.md exakt. Du darfst unmittelbar danach eine deutsche Erklärung in Klammern ergänzen, aber den Namen nicht ersetzen oder abkürzen. Zeige vor Änderungen eine Liste der kanonischen Namen, die du verwenden wirst.

The useful pattern is:

Policy Gate (Prüfstufe für Richtlinien)

The unsafe pattern is:

Politik-Zaun

Ask for a read-only terminology map before asking for edits. Confirm that the map contains the exact tokens from the repository. If it already contains aliases, correct the map and start a fresh implementation turn before any file write.

Add a Deterministic Drift Check

Prompt instructions are probabilistic. A repository check is observable. Start with the exact mistranslations that have actually appeared in your project:

rg -n --glob '*.md' --glob '*.json' --glob '*.yaml' \
  'Politik-Zaun|Richtlinien-Zaun|Vertragsregistrierung|Arbeiter-Pool'

An empty result means those known aliases are absent. A match is a review prompt, not automatic proof of an error: the term may appear in a migration note, test fixture, or article discussing the bug.

Also inspect the actual patch:

git diff --check
git diff --word-diff=plain -- docs/ ARCHITECTURE.md CLAUDE.md

Review these surfaces first:

• architecture and implementation plans;

• memory and session-summary files;

• schema, state, and process names;

• hook and subagent instructions; and

• filenames proposed or created during the task.

If this failure is frequent, put the rg command in an existing documentation or validation script. Keep the forbidden-alias list beside the canonical glossary so the rule has one owner. Do not duplicate the same list in hooks, CI, prompts, and multiple docs.

The agent verification guide explains why an agent's success message is weaker than a command whose result you can inspect.

Recover Without Deleting Project State

When drift has already entered the repository:

1. Stop new edits and save the current diff.

2. Identify the first changed artifact that introduced the alias.

3. Replace only the incorrect terminology while preserving the intended technical change.

4. Check memory, summaries, plans, and subagent output for the same alias.

5. Add the observed alias to the terminology file's forbidden-replacement column.

6. Start a fresh session, ask it to read the glossary, and rerun the deterministic check.

Do not delete all Claude settings, memory, or project documentation. Broad deletion removes evidence and can erase valid context. Do not run a global replacement across source code without reviewing each match; a translated word may be legitimate user-facing copy.

If the changed files are already committed, use a normal review and follow-up commit. Avoid rewriting shared history merely to hide the terminology error.

Verify the Containment Workflow

Use a small, reversible test before trusting the workflow on a long task:

1. Define four canonical names in docs/terminology.md.

2. Ask Claude Code in German to explain the architecture without editing.

3. Ask it to add one paragraph to a temporary documentation file.

4. Run the forbidden-alias search and inspect the diff.

5. Trigger a summary or begin a new session, then repeat the explanation.

Containment is working when the exact canonical tokens survive the explanation, file edit, summary boundary, and fresh session. A single passing run does not prove that the upstream behavior is fixed. Keep the review gate until Anthropic documents a correction and your own repeated test remains clean.

Monitor the Upstream Issue With Aident Loadout

You can monitor the public issue without putting a GitHub token in the prompt. Start with:

Follow https://aident.ai/SETUP.md

Then ask your agent:

Check Aident Loadout authentication and Vault status. Discover the current read-only GitHub issue Action, inspect its schema, and preflight a lookup for anthropics/claude-code issue 85432. If it is valid and free, execute it and return only state, updated time, labels, comment count, title, and canonical URL. Do not comment, react, edit, close, or create anything.

The Aident Loadout first-task guide covers authentication, Vault, discovery, schema inspection, preflight, and execution. The measurable result is one current issue-state record with zero GitHub writes.

Set up Aident Loadout and monitor terminology drift.

Sources

• Claude Code issue 85432: German canonical terminology drift

• Claude Code German glossary

• Claude Code memory documentation

• Claude Code tool reference

• Reddit discussion: Claude Code inventing project jargon

• Reddit discussion: filenames and process terminology

Refresh this guide when issue 85432 changes state, Anthropic adds a terminology-preservation mechanism, or Claude Code changes memory, compaction, or localization behavior.