Mandarin packs more meaning per character than English, so the same brief costs fewer tokens on most tokenizers. Flip this on and every prompt you send to a sub-agent ships in compressed Mandarin while your source-of-truth and all results stay in English. Same instructions, smaller bill.
When to use
- At the start of a session that will spawn sub-agents (Agent/Task tools) more than once or twice.
- Long, instruction-heavy briefs going to multiple agents. The longer the prose, the bigger the cut.
Behavior
- On activation, confirm the mode in one line and keep it on for the rest of the session.
- Before every sub-agent dispatch, translate the instruction prose into clear, dense Simplified Chinese. Keep every constraint, acceptance criterion, and edge case intact.
- Never translate: code, file paths, identifiers, URLs, CLI commands, API and product names, or anything the agent must echo verbatim. Those stay exactly as written.
- Back-check before sending: reread the Mandarin against the English intent. If any clause turns ambiguous in Mandarin, keep that clause in English rather than risk drift.
- Tell every sub-agent to reason in whatever language it likes and RETURN results in English.
- Skip translation when a prompt is mostly code or under about 50 words. Translating those saves nothing.
- Everything the user sees stays in the user's language. Mandarin lives only on the wire between agents.
Output
Dispatches go out in Mandarin; results and all user-facing output come back in English. On the first dispatch, note the rough token savings in one line (typical prose briefs compress 30 to 50 percent).