prompt-factory（提示工厂）

Prompt Factory - World-Class Prompt Powerhouse

Generate production-ready mega-prompts through intelligent questioning and comprehensive domain presets.

⚠️ CRITICAL CONSTRAINTS - READ FIRST

This skill generates PROMPTS only. It does NOT implement the work described in the prompt.

What This Skill DOES:

✅ Generate a comprehensive PROMPT (text document in chosen format) ✅ Ask 5-7 questions to understand requirements (MANDATORY - no skipping) ✅ Validate prompt quality before delivery ✅ Output a SINGLE prompt document with token count ✅ Provide the prompt ready to copy and use elsewhere

What This Skill DOES NOT Do:

❌ Implement the actual work (no code files, no diagrams, no APIs) ❌ Create architectural diagrams or technical implementations ❌ Write actual marketing campaigns or business strategies ❌ Build infrastructure or deploy anything ❌ Execute the prompt after generating it

Expected Workflow:

1. User asks for help creating a prompt
2. Skill MUST ask 5-7 questions (even if context seems obvious)
3. User answers questions with specific details
4. Skill generates ONE comprehensive prompt document
5. Skill announces token count
6. STOP - Do not implement anything from the prompt
7. Ask: "Would you like me to modify the prompt or create a variation?"

Gotchas & Common Pitfalls

1. Skipping Questions

Danger: Agent rushes to prompt generation without asking questions

• Why it breaks: Vague prompts miss key context, leading to poor outputs when user applies them
• Fix: ALWAYS ask 5-7 questions minimum, even for "obvious" requests
• Example: User says "PM prompt for PRD" but doesn't specify domain, constraints, or team context

- Wrong: Jump to generate PM preset - Right: Ask: domain, PRD type, team size, constraints, success criteria

2. Implementing Instead of Generating

Danger: Agent starts coding/building the thing the prompt describes

• Why it breaks: Scope creep; this skill only makes prompts, not implementations
• Fix: Generate the prompt → Stop → Offer variations only
• Example: Prompt describes "REST API for payments"

- Wrong: Start coding the API - Right: Generate prompt → User takes prompt elsewhere to implement

3. Conflating with PROMPTS_FACTORY_PROMPT.md

Danger: User asks "make a prompt system for FinTech" and you generate one mega-prompt

• Why it breaks: That's a meta-prompt task (building a prompt builder), not a single role prompt
• Fix: Redirect to PROMPTS_FACTORY_PROMPT.md for domain-wide builders
• Difference:

- This skill: One prompt for one role (e.g., "FinTech PM") → ~5K tokens - Meta-prompt: Entire FinTech system with 10-20 role presets → generate with PROMPTS_FACTORY_PROMPT.md

4. Token Count Inflation

Danger: Generated prompts exceed optimal ranges (core: 3-6K, advanced: 8-12K)

• Why it breaks: Oversized prompts are harder to use; undersized prompts lack specificity
• Fix: Apply optimization tips from references/prompt-patterns.md
• Monitor: Announce token count; flag if core >8K or advanced >15K

5. Generating Multiple Formats by Default

Danger: Generating all 4 formats when user asked for one

• Why it breaks: Token bloat; user only needs one format
• Fix: Ask for format preference first; only generate all formats if user requests
• Default: XML (optimal for LLM parsing)

6. Weak Question Flow

Danger: Asking 7 generic questions, not 5-7 contextually relevant ones

• Why it breaks: Questions don't validate assumptions; responses lack specificity
• Fix: Use smart adaptation in references/advanced-workflow.md → Skip truly redundant questions, emphasize domain/constraints
• Example: If user specifies "React 18 + TypeScript," skip tech stack question

- Still ask: domain, task, constraints, success criteria (always validate)

Quick Start: Choose Your Path

Path 1: Quick-Start Preset (Fastest)

Use when: You need a prompt for a common role

1. User says: "I need a prompt for [preset name]"
2. Confirm the preset matches their need
3. Customize any variables (optional)
4. Generate → Deliver with token count
5. Ask: "Want a variation or different format?"

Available Presets: 69 across 15 domains (see references/presets.md)

Path 2: Custom Prompt (5-7 Questions - MANDATORY)

Use when: Building a unique prompt from scratch

1. Detect intent from user request
2. MUST ask 5-7 questions with example answers (mandatory - even for "obvious" requests)
3. Apply contextual best practices
4. Validate quality → Deliver with token count
5. Ask: "Want to modify the prompt or try a different format?"

Full 8-step workflow: See references/advanced-workflow.md

Expected Output Formats

Choose one or generate all:

1. XML (default) — Optimal for LLM parsing with structured tags
2. Claude — Claude system prompt format
3. ChatGPT — Custom instructions format
4. Gemini — Google Gemini format
5. All — Generate all 4 formats (use sparingly - token heavy)

See references/examples.md for complete examples of each format.

Generation Modes

1. Core (default) — Prompt + instructions + 2-3 examples (~4-6K tokens)
2. Advanced — Core + testing scenarios + variations + optimization tips (~10K tokens)

Quality Validation (7-Point Gates)

Before delivery, validate:

1. ✓ XML Structure valid (if XML format)
2. ✓ Completeness — All questions incorporated
3. ✓ Token Count optimal (core: 3-6K, advanced: 8-12K)
4. ✓ No Placeholders — All [...] filled
5. ✓ Actionable Workflow — Clear, executable steps
6. ✓ Best Practices applied contextually
7. ✓ Examples present (minimum 2)

If validation fails: Fix before delivery.

Token Count Announcement

After generating, announce:

• "Token Count: ~4,200 tokens (Core mode - within optimal range ✅)"
• "Token Count: ~10,500 tokens (Advanced mode - comprehensive ✅)"
• "Token Count: ~7,800 tokens (Warning: Higher than typical Core mode)"

Reference Files

Read on-demand based on context:

Modern Prompting Techniques (2025/2026)

When generating prompts, incorporate these proven techniques as appropriate:

Chain-of-Thought (CoT)

For reasoning tasks, instruct the model to think step-by-step:

<instructions>
Think through this step-by-step before giving your final answer.
Show your reasoning process explicitly.
</instructions>

Structured Outputs

For prompts that need machine-parseable responses:

<output_format>
Respond ONLY with valid JSON matching this schema:
{"result": string, "confidence": number, "reasoning": string}
</output_format>

Few-Shot Examples

Always include 2-3 high-quality examples for consistent behavior:

<examples>
<example>
<input>Schedule a meeting for next Tuesday</input>
<output>{"action": "create_event", "day": "next_tuesday", "title": "Meeting"}</output>
</example>
</examples>

Role + Context + Task Structure (optimal for Claude)

<role>You are a senior {domain} specialist with 15 years of experience at top-tier firms.</role>
<context>You are helping {user_type} with {specific_situation}.</context>
<task>{specific_instructions}</task>
<constraints>{limitations_and_guardrails}</constraints>
<output_format>{format_specification}</output_format>

Negative Instructions

Explicitly stating what NOT to do reduces unwanted behaviors:

<constraints>
- Do NOT include unsolicited advice beyond the specific question
- Do NOT use jargon without explanation
- Do NOT recommend third-party tools unless asked
</constraints>

Model-Specific Formatting Notes

• Claude: Responds best to XML tags; prefers explicit thinking steps; use <thinking> for CoT
• GPT-4.1: Responds well to markdown; system prompt is highly influential
• Gemini 2.5: Responds well to numbered lists; excels with explicit JSON schema specs

Common Questions

"Do you implement the prompt after generating it?" → No. This skill makes prompts only. Once generated, you take it elsewhere to implement or use with another LLM.

"Can I get a variation of this prompt?" → Yes. Ask for "concise version," "more detailed," "different tone," or "different format." We'll modify the prompt (not reimplement it).

"What if the generated prompt doesn't work?" → We can refine it. Share what didn't work, and we'll adjust specific sections, not rebuild from scratch.

"Why so many questions?" → Good prompts need context. 5-7 questions ensure we capture domain specifics, constraints, and success criteria—making the final prompt much more effective.

Ready to generate a world-class prompt? Let's start with a preset name or custom request.