Gc AI Speak Translator

GC AI Speak Translator

Use this skill to convert the user's natural expression into instructions another AI can execute. Treat this as intent translation, not generic prompt beautification.

Core Workflow

1. Preserve the user's original intent. Do not make the request broader, more polite, more complex, or more ambitious unless the user asks for that.
2. Identify the target AI type: coding AI, research AI, writing AI, image AI, general chat AI, or agent AI.
3. Extract the task:
   • What should the target AI do?
   • What input, file, text, context, or artifact should it use?
   • What should the final output look like?
   • What must it avoid?
4. Separate explicit requirements, reasonable assumptions, and missing information.
5. Ask at most one concise clarification question only when the missing information would materially change the task. Otherwise, state assumptions and continue.
6. Write in the language the target AI should receive. Default to Chinese when the user writes Chinese.

Mode Selection

Choose one mode before writing the prompt:

• coding-ai: For Cursor, Codex, Claude Code, repo edits, debugging, UI polish, tests, or implementation tasks. Include files or areas to inspect, desired behavior, constraints, acceptance criteria, and verification steps.
• research-ai: For reading documents, comparing ideas, searching, analyzing claims, or producing judgment. Include source scope, evidence requirements, uncertainty handling, and output structure.
• writing-ai: For rewriting, editing, article drafting, tone transfer, summarization, or content planning. Include audience, purpose, tone, structure, length, examples to preserve, and things to avoid.
• image-ai: For image generation or visual prompt work. Include subject, scene, composition, style, mood, constraints, and avoid inventing details that change the user's idea.
• general-chat-ai: For ordinary ChatGPT or Claude requests. Include role only when useful, task, input, process constraints, and expected output.
• agent-ai: For autonomous workflows. Include goal, available context, tool permissions, stop conditions, progress reporting, and final deliverable.

If the target AI is not named, infer the mode from the task. If several modes fit, choose the narrowest useful mode and mention the assumption.

Output Levels

Use the lightest output that solves the request:

• Short request: return only 可直接复制版本.
• Medium request: return 原意判断, 给 AI 的任务说明, 约束, and 可直接复制版本.
• Complex or risky request: return the full structure below.

用户原话：
[The user's original request, quoted or lightly summarized.]

原意判断：
[What the user is really asking for.]

目标 AI 类型：
[coding-ai / research-ai / writing-ai / image-ai / general-chat-ai / agent-ai]

给 AI 的任务说明：
[A clear instruction the target AI can follow.]

背景：
[Relevant context from the user's words.]

目标：
[The desired outcome.]

约束：
- [Requirement or boundary.]
- [Requirement or boundary.]

输出格式：
[What the AI should return.]

缺失信息或默认假设：
- [Question or assumption.]

保守翻译版：
[Strictly based on what the user said.]

增强执行版：
[A stronger executable version with reasonable structure added.]

可直接复制版本：
[The best final prompt to paste into another AI.]

Omit empty sections. If the user explicitly asks for only one prompt, provide only 可直接复制版本.

Translation Rules

• Preserve intent; improve executability.
• Keep the user's strength of judgment when it matters. Remove noise, not standards.
• Do not invent domain facts, files, names, deadlines, budgets, user preferences, or business background.
• Turn fuzzy verbs into executable actions: analyze, compare, rewrite, implement, debug, summarize, extract, design, validate, or ask clarifying questions.
• Convert hidden quality bars into checkable criteria when implied:
  • "不要糊弄" -> "指出不确定处，不要编造。"
  • "做得好一点" -> "给出判断依据和可验证步骤。"
  • "别瞎改" -> "保留现有功能和数据结构，只修改指定范围。"
• Treat missing information as a default assumption unless it changes direction, scope, risk, cost, or output.
• For coding tasks, include verification. A prompt that asks for code changes but has no acceptance criteria is incomplete.
• For research tasks, require evidence and uncertainty labels. Do not let the target AI present guesses as facts.
• For creative tasks, preserve the user's core image or tone. Do not overwrite it with generic cinematic language.

Do Not

• Do not add empty prompt boilerplate such as "你是世界顶级专家" unless it genuinely helps the task.
• Do not automatically add "一步一步思考" or chain-of-thought requests.
• Do not make the prompt more complicated just to look professional.
• Do not flatten the user's real standard into polite vagueness.
• Do not translate ordinary language translation requests unless the user means "translate my intent for an AI" rather than "translate into another human language."
• Do not trigger for plain polishing, copyediting, or article writing unless the user specifically wants a prompt or task brief for another AI.

Examples

User: "帮我给 Cursor 说一下：这个页面太乱了，让它改得专业一点，别瞎改功能。"

Useful translation:

可直接复制版本：
请检查当前页面的结构和样式，优化视觉层级、间距、对齐、信息分组和响应式表现，让页面更专业、更易扫描。不要改变现有功能、数据流、路由、交互逻辑或文案含义。完成后请说明你改了哪些文件、为什么这样改，并运行可用的 lint/test/build 或页面验证步骤。

User: "给 Claude 说清楚：读这个文档，别总结废话，告诉我哪里能落地。"

Useful translation:

可直接复制版本：
请阅读我提供的文档，不要做泛泛摘要。请提取其中真正可执行的做法，按「可落地动作」「适用前提」「需要的资源」「潜在风险」「第一步怎么做」输出。遇到文档没有证据支持的部分，请标为不确定，不要自行编造。

Quality Check

Before finalizing, verify that the translated prompt answers:

• What should the target AI do?
• Which AI mode is being addressed?
• What context or input should it use?
• What should it avoid?
• What output format is expected?
• Which assumptions are being made?
• For coding or agent tasks, how will completion be verified?

If any answer is missing and important, add it as a question or assumption instead of silently filling it in.