AgentSkillsCN

youtube-transcript-to-skill

将 YouTube 视频的字幕转换为 Claude Code 技能。提取实用技巧，依据筛选标准进行评估，并以完全透明的方式生成 SKILL.md 文件。

SKILL.md
--- frontmatter
name: youtube-transcript-to-skill
description: Convert a YouTube video transcript into Claude Code skills. Extracts tips, evaluates against selection criteria, and generates SKILL.md files with full transparency.
disable-model-invocation: true
argument-hint: [path/to/transcript.txt]

YouTube Transcript to Skill

Convert a video transcript (or article) about Claude Code / AI coding tools into reusable skills. Extracts discrete tips, evaluates each against a structured selection framework, and generates SKILL.md files for tips that qualify — with full transparency at every step.

Usage

/youtube-transcript-to-skill path/to/transcript.txt

Supported file types: .txt, .md, .pdf

Procedure

Step 1: Load and Validate Transcript

  1. •If $ARGUMENTS is empty, ask the user: "Please provide a file path to the transcript (.txt, .md, or .pdf)."
  2. •Check the file extension:
    • •Supported: .txt, .md, .pdf — proceed
    • •Unsupported: Stop and tell the user: "I can't read .<ext> files directly. Please convert it to .txt, .md, or .pdf first, then re-run the skill."
  3. •Read the file using the Read tool. For large PDFs (10+ pages), read in chunks using the pages parameter.
  4. •Estimate the word count and assess the topic.
  5. •Topic check: If the content is clearly unrelated to Claude Code, AI coding tools, or developer workflows (e.g., a cooking tutorial, a political speech), warn the user:

    "This transcript appears to be about [detected topic], not about Claude Code or AI-assisted development workflows. No skills can be extracted. Would you like me to proceed anyway in case I missed something?" If the user says stop, end here. If they say proceed, continue but expect zero skill-worthy tips.

  6. •Show: "Loaded transcript: ~N words from [filename], appears to be about [topic]."

Step 2: Extract Discrete Tips

  1. •Parse the transcript into individually actionable tips or techniques. A tip is a discrete, specific piece of advice that describes something a developer could DO — not a general observation or opinion.
  2. •Number them sequentially starting from 1.
  3. •For each tip, extract:
    • •Short name: 3-6 word label (e.g., "Plan before coding", "Auto-format with hooks")
    • •Description: 1-2 sentences explaining what the tip recommends
  4. •Show the full numbered list to the user.
  5. •If zero tips were extracted, report:

    "I couldn't extract any discrete, actionable tips from this transcript. The content appears to be [description]. Tips need to be specific techniques or workflows that Claude could execute. No skills can be generated." Write the analysis report (Step 8) explaining why, then stop.

  6. •Ask: "I found N tips. Any to add, remove, or split?"
  7. •Incorporate the user's feedback before proceeding.

Step 3: Load Selection Criteria

Read the selection criteria reference file bundled with this skill:

code
.claude/skills/youtube-transcript-to-skill/selection-criteria.md

This contains the full Gate (G1-G4), Value (V1-V4), and Exclusion (E1-E6) framework used for evaluation.

Step 4: Evaluate Each Tip

For EACH extracted tip (show results as you go, not batched at the end):

  1. •Gate G1 — Automatable: Can Claude execute this programmatically? If no, classify as UI/Shortcut and move on.
  2. •Gate G2 — Recurring: Would someone invoke this more than once? If no, classify as One-time.
  3. •Gate G3 — Not built-in: Does Claude Code already do this? If yes, classify as Built-in.
  4. •Gate G4 — Best mechanism: Is a Skill the right mechanism?
    • •If it should always apply -> classify as CLAUDE.md
    • •If it must run deterministically on events -> classify as Hook
    • •If it requires external service connection -> classify as MCP
  5. •Value check: For tips that pass all gates, check V1-V4. At least one must pass. If none pass, note why.
  6. •Exclusion check: Check E1-E6. If any exclusion applies, classify accordingly.

For each tip, show a brief verdict:

"Tip N: [name] — [PASS/FAIL] — [reason]. Category: [category]"

Step 5: Present Classification Summary

  1. •Show a summary table grouping all tips by their classification category:
code
| Category       | Count | Tip Numbers        |
|----------------|-------|--------------------|
| Skill-worthy   | N     | #3, #7, #12, ...   |
| UI/Shortcut    | N     | #1, #6, ...        |
| Built-in       | N     | #4, ...            |
| One-time       | N     | #2, ...            |
| Philosophy     | N     | #8, ...            |
| CLAUDE.md      | N     | #5, ...            |
| Hook           | N     | #9, ...            |
| MCP            | N     | #11, ...           |
  1. •For each rejected tip, show the specific gate/exclusion that failed and what mechanism would be better.
  2. •If zero tips are skill-worthy, report:

    "I extracted N tips, but none passed the skill selection criteria. Here's why: [summary]. No SKILL.md files will be generated, but the full analysis is saved to [report path]." Write the analysis report (Step 8), then stop.

  3. •Ask: "Any disagreements with this classification?"
  4. •Incorporate feedback before proceeding.

Step 6: Group Tips into Skills

  1. •Examine the skill-worthy tips for natural groupings:
    • •Shared domain: Tips about the same subsystem or feature area
    • •Sequential workflow: Tips that form a natural ordered sequence
    • •Setup + usage: Configuration tip paired with a usage tip
  2. •For each proposed skill, define:
    • •Skill name (kebab-case)
    • •Description (one sentence)
    • •Source tips (which tip numbers it covers)
    • •Grouping rationale (why these tips belong together)
  3. •Prefer fewer, more capable skills over many tiny ones. A skill with 3-6 steps is ideal.
  4. •Show the grouping to the user.
  5. •Ask: "Should I adjust any groupings?"
  6. •Incorporate feedback before proceeding.

Step 7: Generate Skills

For each proposed skill:

  1. •Generate a complete SKILL.md following project conventions:
    • •YAML frontmatter with name, description, disable-model-invocation: true, and argument-hint (if the skill takes input)
    • •H1 title and brief description
    • •## Usage section with invocation syntax
    • •## Procedure section with numbered steps (### Step N: Name)
    • •Each step should have clear, specific instructions Claude can follow
  2. •Show the full generated SKILL.md to the user.
  3. •After showing all generated skills, ask: "Write these skills to disk?"
  4. •If the user wants changes, revise and re-show before writing.

Step 8: Write Output

  1. •Write SKILL.md files to the current project's .claude/skills/<skill-name>/SKILL.md.
    • •NEVER write to ~/.claude/skills/ (user-global scope).
    • •Create the directory if it doesn't exist.
  2. •Write the analysis report to the current project's docs/<video-slug>-analysis.md. Derive the slug from the video title or filename (kebab-case, lowercase). The report contains:
markdown
# Transcript Analysis: [Title]

**Source**: [filename]
**Author**: [if known]
**Topic**: [detected topic]
**Tips extracted**: N
**Skills generated**: M

## Extracted Tips

[Numbered list of all tips with short name and description]

## Evaluation Results

[Per-tip evaluation showing gates, values, exclusions, and verdict]

## Classification Summary

[Summary table by category]

## Skill Grouping

[Which tips combined into which skills, with rationale]

## Generated Skills

[List of skills created with file paths]

## Non-Skill Recommendations

[Tips better served by other mechanisms, with specific guidance]
  1. •Show a final summary:

    "Done. Extracted N tips, created M skills, redirected K tips to other mechanisms. Analysis report: docs/[slug]-analysis.md. Skills installed to .claude/skills/ (project-local only)."