skill-to-prompt（提示技巧）

Skill to Prompt Converter

Converts a Claude skill (SKILL.md + references/) into ChatGPT Project format: one prompt instruction document (.docx, under 8,000 characters) and one knowledge file document (.docx).

What This Produces

For each conversion:

1. Prompt Instructions (.docx) — The text that gets pasted into ChatGPT's "Instructions" field. Under 8,000 characters. Structured using the 9-component prompt formula.
2. Knowledge File (1.docx file) — Uploaded to the ChatGPT Project as reference material. Contains all detailed workflows, templates, frameworks, examples, and reference content consolidated into a single document.

All.docx files follow standard formatting: Arial font, 10pt normal text, H1=20pt, H2=16pt, H3=14pt, H4=12pt, 1.15 line spacing.

Conversion Workflow

Step 1: Locate and Read Source Skill

If user provides a skill name (e.g., "copywriting"):

• Look in ~/.claude/skills/[skill-name]/SKILL.md

If user provides a path, use it directly.

Read everything:

• SKILL.md (full content)
• All files in references/ folder
• Note any scripts/ or assets/ folders (these need special handling)

Step 2: Analyze Skill Complexity

Count total source content to understand how much compression is needed for the single knowledge file:

Regardless of complexity, always produce exactly 1 knowledge file.

Step 3: Map Content to the 9-Component Prompt Formula

Read references/prompt-structure-guide.md for the full mapping guide.

Extract from the Claude skill and map to ChatGPT instruction components:

Step 4: Split Content Between Instructions and Knowledge Files

Apply the priority system to stay under 8,000 characters:

P1 — Always in the prompt instructions:

• Role statement (who the AI is, what it specializes in)
• Task overview (primary task + condensed workflow steps)
• Goal (desired outcome)
• Critical constraints (never/always rules)
• Knowledge file references (what each file contains, when to consult it)

P2 — Include in instructions if character budget allows:

• 1-2 key examples (keep short, move full examples to knowledge files)
• Input variables (what the user must provide before starting)
• Output formatting specs
• Style/tone guidance

P3 — Always in knowledge files:

• Full detailed workflow with multi-paragraph explanations
• Complete example library
• Templates and frameworks (from SKILL.md body and references/)
• All reference material (from references/ folder)
• Edge cases and advanced scenarios
• Decision trees and detailed tables
• Script code (from scripts/ folder, documented as reference)

Character budget: Target 7,500 characters for the instruction text (500-char safety buffer).

Step 5: Write Prompt Instructions

Draft the instruction text following this structure:

Every prompt instruction document must include all 9 headings below, in this exact order. No headings may be omitted or renamed.

## Role

You are [expertise extracted from skill description and intro].
You specialize in [specific domain/capabilities].

## Task

Your primary task is to [main objective from the skill].

Follow this workflow:
1. [Condensed step from skill workflow]
2. [Condensed step]
3. [Condensed step]
...

## Goal

[Desired outcome extracted from the skill's purpose/output sections]

## Audience

[Target audience. If the skill does not specify an audience, write: "General users seeking [skill domain] assistance. Adjust complexity based on the user's apparent experience level."]

## Style/Tone

[Voice/tone rules extracted from the skill. If the skill does not specify tone, write: "Professional and clear. Match the register of the target audience. Avoid jargon unless the audience expects it."]

Always include these writing rules regardless of skill-specific tone:
- Avoid dashes unless absolutely necessary. Prefer commas, periods, or conjunctions to keep the tone natural.
- Use en dashes (with spaces, e.g. word – word) or em dashes (without spaces, e.g. word—word) only when unavoidable, ideally no more than once per complete piece of output.

## Constraints

- [Never/always rule]
- [Character/word limits]
- [Format restrictions]
- [Quality standards]

## Inputs/Placeholders

Before starting, ask the user for:
1. [Required input from the skill's context-gathering steps]
2. [Required input]

## Example Output

[1-2 short examples showing expected output structure. Keep brief.]

## Formatting Instructions

[Output structure and formatting rules from the skill]

After the 9 required sections, always append a Knowledge Files section:

## Knowledge Files

This project includes 1 knowledge file. Consult it as needed:

1. **[SkillName]_Reference_Guide.docx** — [What it contains and when to reference it]

After drafting, count characters. If over 7,500:

• Cut examples to bare minimum (but keep the Example Output heading with at least a brief example)
• Compress workflow steps further
• Shorten Audience and Style/Tone content (but never remove the headings)

If over 8,000 after compression: shorten section content further, but all 9 headings must remain. Move detailed explanations to the knowledge file and replace with concise summaries under each heading.

Step 6: Organize Knowledge File Content

Always produce exactly 1 knowledge file named [SkillName]_Reference_Guide.docx.

Consolidate all P3 content into this single document, organized with clear heading hierarchy:

[SkillName]_Reference_Guide.docx containing:

• Full detailed workflow with complete explanations
• All examples and worked examples
• Templates and frameworks (from SKILL.md body and references/ folder)
• All reference material (from references/ folder)
• Guidelines, best practices, and edge cases
• Decision trees and detailed tables
• Script documentation (from scripts/ folder, noted as non-executable reference)

Organization within the single file:

• Use H1 for major sections (Workflows, Templates, Examples, Reference Material)
• Use H2-H3 for subsections within each major section
• Use page breaks between major sections for readability
• For complex skills with heavy reference content, prioritize frameworks, worked examples, and actionable content over verbose explanations

Step 7: Generate JSON Data

Structure all content as a JSON object matching the docx-generator input format.

The JSON structure:

{
  "skillName": "Skill Display Name",
  "originalSkillName": "skill-name",
  "version": "1.0.0",
  "outputDir": "~/Downloads/",
  "instructionDoc": {
    "filename": "SkillName_Prompt_Instructions.docx",
    "title": "Skill Name — ChatGPT Project Instructions",
    "characterCount": 7234,
    "sections": [
      { "type": "heading", "level": 1, "text": "Section Title" },
      { "type": "paragraph", "text": "Regular paragraph text." },
      { "type": "bold-paragraph", "label": "Label:", "text": "Text after label" },
      { "type": "bullet", "text": "Bullet point", "level": 0 },
      { "type": "numbered", "text": "Numbered item", "number": 1 },
      { "type": "table", "headers": ["H1","H2"], "rows": [["a","b"]], "columnPcts": [30,70] },
      { "type": "divider" },
      { "type": "spacer", "size": 200 },
      { "type": "meta", "text": "Metadata or footnote text" }
    ]
  },
  "knowledgeFiles": [
    {
      "filename": "SkillName_Reference_Guide.docx",
      "title": "Skill Name Reference Guide",
      "sections": [...]
    }
  ]
}

Write this JSON to a temporary file in the current directory: [skill-name]-conversion.json

Step 8: Generate.docx Files

Run the generator:

NODE_PATH=~/Documents/node_modules node ~/.claude/skills/skill-to-prompt/references/docx-generator.js [skill-name]-conversion.json

This produces all.docx files in the ~/Downloads/ folder.

If the docx package is not available, install it first:

cd ~/Documents && npm install docx && cd -

Fallback: If.docx generation fails entirely, output the content as formatted markdown that the user can paste into Google Docs or Word. Apply the same heading hierarchy and section structure.

Step 9: Validate and Present Results

Check and report:

• Instruction character count (must be under 8,000)
• All.docx files generated successfully
• Knowledge files contain complete referenced content
• Cross-references between instructions and knowledge files are clear

Present to the user:

Conversion complete: [skill-name] → ChatGPT Project

Output saved to ~/Downloads/

Prompt Instructions:
  [SkillName]_Prompt_Instructions.docx (X,XXX characters)

Knowledge File:
  [SkillName]_Reference_Guide.docx — [brief description]

Setup in ChatGPT:
  1. Create a new Project in ChatGPT
  2. Paste the instruction text from the Instructions .docx into the project's Instructions field
  3. Upload the knowledge file to the project
  4. Test with a sample prompt

After presenting, clean up the temporary JSON file:

rm [skill-name]-conversion.json

Conversion Rules

What Transfers Directly

• Role and purpose definitions
• Workflow steps (condensed)
• Constraints and rules
• Output format specifications
• Examples and templates
• Reference content

What Requires Rewriting

• Claude tool references (Read, Grep, Glob, Bash, Edit, Write) → Rewrite as natural language instructions. Example: "Read the user's file" becomes "Ask the user to paste or upload the content."
• Dynamic commands (!command syntax) → Not available in ChatGPT. Document as manual steps.
• File system operations → Replace with "ask the user to provide" patterns.
• Progressive disclosure (loading references on demand) → All content must be in knowledge files upfront.

What Does Not Transfer

• Scripts (scripts/ folder) → Document the script's purpose in the knowledge file. Include the code as reference if useful, but note it cannot execute in ChatGPT.
• Assets (assets/ folder) → Note in the output: "These assets must be manually transferred if needed."
• Cross-skill references → Flag as dependencies: "This skill references [other-skill]. Consider converting that skill separately."
• MCP server integrations → Not available in ChatGPT Projects.
• Frontmatter fields (version, argument-hint, context, agent, allowed-tools) → These are Claude-specific metadata. Do not include in ChatGPT instructions.

Output Naming Convention

[SkillName]_Prompt_Instructions.docx
[SkillName]_Reference_Guide.docx

Where [SkillName] is the skill name in Title Case with spaces replaced by underscores.

Examples:

• copywriting → Copywriting_Prompt_Instructions.docx
• digital-product-video-ads → Digital_Product_Video_Ads_Prompt_Instructions.docx
• page-cro → Page_CRO_Prompt_Instructions.docx

Document Formatting Standards

All.docx output follows these standards:

Additional rules:

• Headings are styled by size only. No extra bold on headings.
• All text and headings must be default black. No additional colors on any text or headings.
• Add space after each paragraph.
• Tables use alternating row shading for readability.
• When generating.docx, explicitly set Arial font, heading sizes, black color (no theme colors), and 1.15 line spacing on every element. Never rely on Word default styles.

References

• references/prompt-structure-guide.md — 9-component prompt formula, character budget allocation, compression techniques
• references/docx-generator.js — Node.js script for generating.docx files from JSON data
• references/AI Prompting Techniques.pdf — Source material for the prompt structure framework

Related Skills

• agent-skill-builder — The reverse direction: converts ChatGPT Projects into Claude skills
• skill-creator — Guide for creating new Claude skills from scratch