Troubleshooting
Agent Replies in the Wrong Language?
Fix language drift fast when your agent keeps switching languages mid-task.
A recurring friction point: you ask in one language, the agent answers in another, or flips language halfway through a workflow. This is usually a context and instruction-priority issue, not a broken model.
Root cause: language drift usually comes from mixed-language context, copied prompts in a different language, or missing output constraints.
Fast 4-step recovery
1) Set one explicit language lock
Start your next message with a direct rule: “Reply only in [LANGUAGE] unless I explicitly ask to switch.”
2) Reset context for clean behavior
If the thread is mixed-language, open a fresh session and paste a compact handoff plus language lock.
3) Add output format requirements
Specify tone + format in the same instruction block. Example: bullet list, no translation, keep technical terms in English only.
4) Require self-check before final output
Have the agent confirm the output language before sending the final answer.
Paste-ready language lock prompt
For this thread, reply ONLY in [LANGUAGE].
If I send text in another language, keep your reply in [LANGUAGE] unless I explicitly say: "switch language".
Before your final answer, run a quick check and confirm: "Language check passed: [LANGUAGE]".
Then provide the answer.
High-friction mistakes to avoid
- Mixed-language examples: if your sample prompt is in another language, the agent may mirror it.
- No format constraints: vague asks invite style/language drift.
- Switching surfaces: dashboard/DM/server threads can have different context history.
- Long pasted transcripts: old multilingual text can override your current preference.
Important: if you actually need bilingual output, ask explicitly for structure (e.g., “Spanish first, English translation second”). Don’t assume the agent will infer your format.
Known-good bilingual template
Output format:
1) Spanish version (plain language)
2) English translation
3) Keep product/tool names untranslated
4) Keep each section under 120 words
When to escalate in #help
If language drift continues after a fresh session + lock prompt, share:
- The exact language-lock prompt you used
- Which surface you tested (dashboard/DM/server)
- One short example of expected vs actual output
- Whether behavior changed after a fresh session
Bottom line: treat language like a hard output constraint, not a soft preference. One explicit language lock + fresh context resolves most cases quickly.