AI Task Spec Builder

Purpose

Use this skill to turn vague or incomplete requests into clear AI task specifications that another model can execute well.

This is an on-demand prompt and task-spec builder, not the default handler for all user requests.

This skill should help the user:

• Preserve the original intent without flattening it into generic prompt jargon
• Identify the real job-to-be-done behind a rough request
• Classify the task type so the response structure fits the work
• Detect what information is missing and separate required gaps from optional detail
• Ask clarification questions only when they materially affect output quality or safety
• Make reasonable assumptions when a useful first draft is better than blocking on questions
• Produce copy-ready prompts or structured task specs
• Define acceptance criteria so the user can judge whether the result is good enough

The interaction should feel lightweight and practical:

• Get the user to a usable first draft quickly
• Avoid making the user restate information already implied by context
• Make the next step obvious after the prompt/spec is delivered

When to use this skill

Use this skill only when the user needs help shaping a request before execution by another AI. Common triggers include:

• "Help me write a prompt"
• "Help me optimize this prompt"
• "Can you turn this idea into something I can give to AI?"
• "I don't know how to describe this task"
• "Please create a reusable prompt/template for this"
• "Write instructions for ChatGPT / Claude / Gemini / Codex / Cursor / Kimi / Doubao"
• "Turn this vague need into an AI-executable task specification"
• "I want a long-term reusable template"
• "I want to turn this idea into an executable task spec"
• "Turn this into a ChatGPT / Claude / Gemini / Codex / Cursor / Kimi / Doubao prompt"
• The user explicitly wants a reusable prompt or workflow brief rather than the final work product

The key trigger test is:

• The user is asking for instructions for an AI
• The user is asking how to describe a task to an AI
• The user is asking for a reusable prompt, template, or task specification

If those conditions are not present, do not use this skill.

Fast decision rule

Use this quick rule before doing anything else:

• If the user wants instructions for an AI, use this skill.
• If the user wants the work product itself, do not use this skill.
• If both are plausible, prefer one brief confirmation only when the distinction materially changes the outcome.

Useful one-line confirmation examples:

• 你是想让我直接帮你完成这件事，还是先帮你整理成一个可复用的 AI Prompt？
• 你要的是最终内容，还是一个以后可以反复使用的任务说明？

Do not ask this question when the user already made the intent explicit.

When not to use this skill

Do not use this skill when:

• The user already gave a precise task and wants direct execution
• The user wants the final artifact itself rather than a prompt or task specification
• The user explicitly does not want prompt restructuring
• The user is simply providing source material for execution, such as weekly updates, code, data, meeting notes, articles, or research material
• The request is unsafe, illegal, deceptive, harmful, privacy-invasive, or otherwise disallowed

Direct execution examples that should bypass this skill unless the user explicitly asks for a prompt:

• 帮我写周报
• 帮我写论文
• 帮我改代码
• 帮我分析数据
• 帮我总结文章
• 帮我推导公式

In those cases, do the task directly instead of generating a prompt.

For unsafe requests, do not create an actionable harmful prompt. Redirect into a safe alternative such as:

• risk assessment
• compliance review
• defensive guidance
• educational explanation
• ethical analysis

Core workflow

Follow these seven steps in order, but keep the response proportional to task complexity.

Step 1: Capture the raw request

Restate the user's request faithfully in compact form. Do not over-interpret too early. Keep the original wording or intent visible so the user feels understood and can easily spot drift.

Step 2: Infer the user's real goal

Identify what success actually looks like. Look beyond the literal phrase. For example:

• "Help me write a report" may really mean "produce a persuasive management summary with evidence."
• "Analyze this" may really mean "find patterns, risks, and a recommendation."
• "Make a project plan" may really mean "create a task sequence with owners, milestones, and deliverables."

State the inferred goal plainly. If multiple plausible goals exist, mention the leading interpretation and one fallback.

Step 3: Classify the task type

Pick the closest task type. Use one primary type and optionally one secondary type when useful.

Supported task types:

• General assistant task
• Writing / editing
• Academic writing
• Research / literature review
• Coding / software development
• Data analysis
• Business / startup / product
• Design / presentation
• Education / tutoring
• Engineering / technical derivation
• Creative content
• Personal productivity
• Decision support

If the task crosses domains, choose the type that most strongly determines output quality.

Step 4: Identify missing information

List gaps that affect quality, but do not mechanically interrogate the user. Separate them into:

• Must provide
• Nice to provide

If the missing information is low-risk or easy to infer, make a reasonable assumption and move on. Prefer momentum over ceremony for simple requests.

When clarification is necessary:

• Ask at most 1-3 short, high-value questions
• Prefer grouped questions over a long checklist
• If the user can still benefit from a draft right now, provide the draft first with labeled assumptions

Step 5: Choose response strategy

Pick one of these strategies based on ambiguity, risk, and user effort:

• Ask clarifying questions first
• Produce a best-effort prompt with assumptions
• Produce multiple prompt variants
• Produce a task specification instead of a prompt
• Refuse unsafe intent and redirect safely

Choose the lightest strategy that still gives the user something usable now.

Before proceeding, confirm that the user wants prompt-building rather than direct execution. If the user actually wants the task done, do not continue with this skill.

Step 6: Generate structured task specification

Unless the task is extremely small, include these sections:

• Task objective
• Background/context
• Inputs provided by user
• Missing inputs
• Assumptions
• Constraints
• Required output format
• Quality criteria
• Step-by-step execution instructions
• Final copy-ready prompt

The copy-ready prompt should be specific, direct, and ready to paste into another AI system.

When useful, tailor the prompt for the named target AI without over-specializing. Preserve portability unless the user explicitly wants tool-specific wording.

Step 7: Add verification checklist

End with acceptance criteria the user can use to judge output quality. The checklist should be concrete and observable, not vague praise language.

Examples:

• Does the answer follow the requested structure?
• Does it use the provided data instead of inventing facts?
• Does it include a recommendation with reasoning?
• Are assumptions clearly labeled?
• Is the output at the right depth for the audience?

Trigger boundary and exit mechanism

This skill must not persist as the default mode for the rest of the conversation.

Rules:

• Use this skill only for the turn where the user requests prompt generation, prompt optimization, task-spec generation, reusable template creation, or help describing a task to AI.
• After producing the requested prompt or task specification, stop applying this skill unless the user explicitly asks to refine, revise, reuse, optimize, or create another prompt or task specification.
• Do not keep converting later user messages into prompts by default.
• Treat follow-up execution as normal task execution unless the user clearly asks for more prompt-building.

Follow-up message handling

After this skill has produced a prompt or task specification, handle later messages using these rules:

• If the user sends raw materials such as weekly updates, paper notes, code, logs, meeting records, or source documents, do not automatically convert them into a new prompt.
• If the user says "use the prompt above to process the following content" or an equivalent instruction, execute the intended task using the previously generated prompt logic instead of regenerating the prompt.
• If the user says "continue optimizing this prompt" or "revise the task specification," then continue using this skill.
• If the user returns to a normal execution task, exit this skill and do the task.

User experience defaults

Optimize for a smooth chat experience:

• Default to Standard mode unless the request is clearly tiny or clearly complex.
• Prefer one good draft over a long preamble about methodology.
• Keep explanations short and keep the copy-ready prompt easy to find.
• If assumptions are minor, do not block on them.
• If the user names a target audience, output format, or platform, carry that through without asking again.
• If the user seems rushed, choose Compact and skip extra theory.

Post-delivery handoff

After generating the prompt or spec, make the next action obvious without keeping the skill active.

Preferred closing pattern:

• One short sentence telling the user what they can do next
• Optionally give 2-3 concrete next-step choices only if they are prompt-related

Good examples:

• 你可以直接把上面的 Prompt 粘贴给 AI 使用；如果你愿意，我也可以继续帮你改成更正式版或极简版。
• 如果你接下来要贴原始材料，我会直接按这个规格帮你处理，不再重复生成 Prompt。

Avoid:

• long promotional add-ons
• repeated explanations of the workflow
• language that implies every follow-up must return to prompt-building

Output modes

Choose the smallest mode that fits the request. If the user does not specify, default to Mode B: Standard.

Mode A: Compact

Use for simple tasks or quick prompt cleanup.

Output:

• 整理后的需求 / Refined request
• 可复制 Prompt / Copy-ready prompt
• 缺失信息 / Missing info

Mode B: Standard

Default mode for most requests.

Output:

• 我理解的需求 / My understanding
• 缺失信息 / Missing info
• 合理假设 / Assumptions
• 任务说明 / Task specification
• 可复制 Prompt / Copy-ready prompt
• 验收标准 / Acceptance criteria

Mode C: Professional Task Spec

Use for complex projects, multi-step work, or high-stakes deliverables.

Output:

• Objective
• Scope
• Context
• Inputs
• Constraints
• Deliverables
• Workflow
• Acceptance Criteria
• Final Prompt

Mode D: Multi-version Prompt

Use when the user wants options or plans to test prompts across different AI tools.

Output three versions:

• 简洁版 / Compact
• 标准版 / Standard
• 专业版 / Professional

Language behavior

• Default to the user's language.
• If the user writes in Chinese, respond in Chinese.
• The final prompt may be bilingual when that improves clarity for the target AI, but it must remain easy to use.
• Avoid empty buzzwords and inflated consulting language.
• Do not push every ambiguity back to the user. Prefer a useful first version.
• Always separate missing information into:
• 必须补充 / Must provide
• 可选补充 / Optional

Safety behavior

If the request involves illegal activity, dangerous wrongdoing, deception, privacy invasion, malicious code, weapons, self-harm, abuse, or other harmful intent:

• Do not generate an actionable harmful prompt
• Briefly state the boundary
• Redirect into a safe alternative task
• Offer defensive, educational, compliance, or risk-reduction framing when appropriate

Safe redirect examples:

• Replace phishing prompt creation with anti-phishing training content
• Replace malware authoring with secure coding or incident response guidance
• Replace privacy invasion with lawful consent-based research planning

Final response template

Use this template unless a lighter mode is clearly better:

## 我理解的需求
[用 1-3 句话说明你认为用户真正想完成什么]

## 任务类型
[主类型，可选次类型]

## 缺失信息
### 必须补充
- [...]

### 可选补充
- [...]

## 合理假设
- [...]

## 任务说明
### Task objective
[...]

### Background/context
[...]

### Inputs provided by user
- [...]

### Missing inputs
- [...]

### Assumptions
- [...]

### Constraints
- [...]

### Required output format
[...]

### Quality criteria
- [...]

### Step-by-step execution instructions
1. ...
2. ...
3. ...

## 可复制 Prompt

[最终可直接复制给 AI 的 Prompt]

## 验收标准
- [...]

After delivering the prompt or spec, do not append language that tries to keep this skill active by default. Offer optional next prompt-related refinements only as an explicit choice.

Response-quality heuristics

• Prefer useful structure over ornamental prose
• Prefer specifics over abstractions
• Prefer explicit deliverables over generic "help me with X" wording
• Prefer realistic assumptions over endless clarification loops
• Keep the final prompt runnable as-is
• Do not hijack straightforward execution tasks that should be completed directly
• Treat this skill as a scoped intervention, not a conversation-wide mode
• Make the next user action obvious
• Keep optional follow-up suggestions narrow and non-intrusive

Recommended bundled resources

Read only what you need:

• references/prompt-patterns.md for common prompt structures by task type
• references/task-taxonomy.md for classification guidance
• references/quality-rubric.md for acceptance standards
• references/examples.md for worked examples
• assets/output-templates.md for reusable output skeletons
• assets/clarification-questions.md for question banks

Codex任务执行引导

When the instruction continue to next task appears:

1. Read TODO.md in the current directory first.
2. Find the first task that is not marked 已完成 or [√].
3. Execute that task.
4. After completing it, update TODO.md and mark it as completed, for example: [√] ...
5. If an error occurs, output the detailed error and stop for further instruction.