minimax-prompting (v1, Phase 5 finalization)

Guidance for Claude when composing a prompt to send to Mini-Agent via minimax-companion.mjs. Not user-facing.

Scope

This skill guides prompt construction for /minimax:ask, /minimax:review, /minimax:rescue, /minimax:adversarial-review. v1 reflects what was actually validated through Phase 1-5 smoke tests against MiniMax-M2 7B / Coding-Plan endpoints.

Universal rules

Output contract first. State the expected output format in the first paragraph. For JSON: explicitly say "Return ONLY a JSON object matching this schema. No prose before or after. No markdown code fence." Echo the schema as a fenced JSON block immediately after.
Context in labeled blocks. Wrap code/diff/docs in labeled blocks (### Diff to review / ### Files under investigation). Do not interleave instructions and content.
Language parity. MiniMax-M2's Chinese reasoning is strong; keep instruction language aligned with user prompt language. Do not force English on Chinese prompts. The output schema enums (severity / verdict) MUST stay English even when surrounding prose is Chinese — explicitly call this out in the prompt (see references/prompt-blocks.md output-contract-bilingual block).
Stance prompts are single-stance. For /minimax:adversarial-review, do NOT mix red and blue stance instructions in one prompt — even if asked nicely, the model biases toward whichever stance appears last. Use two independent spawns (Phase 5 architecture), each with one stance constant from minimax.mjs (RED_STANCE_INSTRUCTION / BLUE_STANCE_INSTRUCTION).