minimax-rules

MiniMax MCP usage boundary rules

Capability list (what MiniMax can do)

Tool	Input → Output	Do I have an alternative
`minimax_text_chat`	Text → Text (translation/polishing/writing/analysis)	✅ DeepSeek can also do it, but MiniMax saves Tokens
`minimax_image_generate`	Text → Image	❌ No alternative, must use
`minimax_video_generate`	Text/Image → Video	❌ No alternative, must use
`minimax_music_generate`	Text → Music	❌ No alternative, must use
`minimax_speech_synthesize`	Text → Speech	❌ No alternative, must use
`minimax_speech_voices`	Query available voices	❌ No alternative
`minimax_vision_describe`	Image → Text description	❌ No alternative, must use
`minimax_search`	Keywords → Search results	✅ web_search exists, but MiniMax has ample quota so prioritize using it

MiniMax's hard limits (things it absolutely cannot do)

❌ Read files —— it can only receive text input, cannot access your disk
❌ Execute commands —— it has no shell or code execution capability whatsoever
❌ Remember context —— each call is stateless, it does not remember what was said in the previous round
❌ Multi-step reasoning —— a single API call, it cannot "first do this then do that"
❌ Call other tools —— it has no MCP/toolchain permissions

Decision tree: who should this task be given to?

用户请求到达
│
├─ 是图片/视频/音乐/语音/视觉/搜索？
│   └─ YES → 无条件用 MiniMax MCP（我没有这些能力）
│
├─ 需要读文件 / 执行命令 / 跨文件分析？
│   └─ YES → 我自己来，❌ 不传给 MiniMax
│
├─ 是纯粹的文本→文本任务（翻译、润色、改写、写文案、写诗、写邮件）？
│   └─ YES → minimax_text_chat（省 DeepSeek Token）
│
├─ 是"让 MiniMax 也看看" / "第二意见"？
│   └─ YES → minimax_text_chat（输入我已有的分析结果或代码片段）
│
├─ 是代码片段级别的简单分析（单函数、单算法、一个 bug 模式）？
│   └─ YES → 可以问 MiniMax 作为补充（不替代我的主分析）
│
└─ 以上都不匹配 → 我自己处理
    ├─ 代码架构 / 重构方案
    ├─ 安全审查 / 漏洞分析
    ├─ 调试 / 排错
    ├─ 数学 / 逻辑推理
    ├─ 多步骤调研
    └─ 需要项目上下文的任务

Chained scheduling rules

The following scenarios automatically orchestrate multi-tool chains, without the need for step-by-step user instructions:

Scenario 1: View image + analyze

用户："这张代码截图有什么问题？"
  step1 → minimax_vision_describe(image)       # 提取图片中的文字
  step2 → 我自己分析提取出的代码                  # 做专业判断
  回复 → 我的分析结果（如有需要，附 MiniMax 的第二意见）

Scenario 2: Search + summarize

用户："搜一下最新的 XXX 并总结"
  step1 → minimax_search(query)                # 获取搜索结果
  step2 → minimax_text_chat("总结这些结果")      # 让 MiniMax 总结
  回复 → MiniMax 的总结

Scenario 3: Translate + polish

用户："把这段话翻译成英文，语气正式一点"
  step1 → minimax_text_chat("翻译并润色...")     # 一步完成
  回复 → 翻译结果

Counterexamples: do not use MiniMax in these situations

Incorrect usage	Why it won't work	Correct approach
"Help me search the project for the usage of XXX"	MiniMax cannot read files	I grep/search myself
"How should this cross-file architecture be changed"	MiniMax cannot see other files	I analyze it myself
"Help me debug this error"	MiniMax has no context and no execution capability	I look at the logs myself
"Analyze the security vulnerabilities of the entire project"	Requires multi-file scanning + professional judgment	I audit it myself
"Help me integrate this piece of code into the project"	Requires modifying files and understanding the project structure	I do it myself

Attitude toward Token usage

User's MiniMax Token Plan quota (domestic version):

Text: 1,500/cycle (ample)
Image: 50/cycle
Music: 100/cycle
Speech: 4,000/cycle
Vision/Search: 150 each/cycle

Strategy: text tasks that can be handled by MiniMax do not consume the DeepSeek reasoning budget. When encountering tasks such as translation, polishing, and copywriting, call MiniMax first by default.