Imported from cdeistopened/skill-stack (public/skills/skill-extractor/SKILL.md). Install upstream with npx skills add cdeistopened/skill-stack --skill skill-extractor. Copyright stays with the author.

Skill Extractor

Turn raw source material into a list of buildable Claude Code skills. Not a knowledge base — skills are workflows Claude can execute, not facts Claude should know.

The Core Distinction

Skill = a repeatable workflow with inputs, steps, and outputs that Claude executes inside a coding environment. Writing a YouTube script. Running an SEO audit. Drafting a newsletter. Creating ad copy.

Not a skill = static knowledge, reference material, domain expertise. How to raise meat birds. The history of fasting. Nutritional science. These become wiki articles, books, or reference docs — not skills.

The test: Can Claude do this thing, repeatedly, with different inputs, and produce a useful output? If yes → skill. If it's something Claude knows and references → not a skill.

Skill Taxonomy

Score each candidate against these six hallmarks. A strong skill exhibits 2+ of these.

1. Repeated SOP / Workflow

Something done regularly as part of a content or business process. The more often it's done, the higher the leverage.

Examples: Newsletter drafting, social post creation, podcast show notes, SEO content briefs, weekly reporting

2. Disposable One-Shot

Valuable but infrequent. Done once per project or client. Still worth codifying because it saves hours when needed — and becomes an agency offering if done for others.

Examples: Writing a LinkedIn bio, Amazon category research, brand identity creation, book launch checklist, landing page copy

3. Specific Knowledge Applied as Process

Domain expertise distilled into a repeatable method — not raw knowledge, but knowledge operationalized into steps Claude can follow.

Examples: Kallaway's "7 Lego Bricks" for short-form → YouTube scriptwriting skill. Dan Koe's "one-person business" model → content strategy skill. NOT "the history of short-form video" (that's reference).

4. Step-by-Step Process

A clear sequence where order matters. Often teachable, often already described as numbered steps in the source material.

Examples: Video editing workflow, podcast production pipeline, email sequence writing, ad creative testing process

5. Example-Driven

Skills where having 3-5 concrete examples dramatically improves output quality. The examples ARE the skill — they teach by pattern, not by instruction.

Examples: Voice/style skills (Trung Phan, Tyler Cowen), ad creative frameworks, hook writing, cold open creation

6. Tool-Augmented

Skills that benefit from MCP servers, API calls, CLI tools, or other integrations. These extend Claude's autonomy — the model can work longer without human input.

Examples: SEO keyword research (DataForSEO), social media scraping (Apify), RSS curation, image generation (Gemini), video processing (ffmpeg)

Available Tool Ecosystem

When evaluating whether a skill candidate is tool-augmented, consider what's currently available:

MCP Servers (live in this workspace):

• Slack (channels, messages, search)
• Google Drive (docs, sheets, slides, folders)
• Google Calendar (events, scheduling, freebusy)
• Apify (web scraping actors — YouTube, Twitter, any site)
• Notion (pages, databases, search)
• Video-audio (trim, convert, subtitles, overlays, concatenate)
• Context7 (library documentation lookup)

CLI Tools:

• ffmpeg (video/audio processing)
• yt-dlp (YouTube downloading)
• Whisper (transcription)
• qmd (local markdown search / RAG)
• gh (GitHub CLI)
• Bun/Node (JS execution)
• Python (scripting, data processing)

APIs (via scripts):

• DataForSEO (keyword research, SERP, rankings)
• Gemini (deep research, image generation, large-context writing)
• ElevenLabs (voice cloning)
• HubSpot (email marketing)
• Webflow (CMS publishing)

When a source mentions a workflow that could be automated with these tools but the creator does it manually, that's a high-value skill candidate.

Extraction Workflow

Phase 1: Ingest and Scan

Input: Raw source material — transcripts, blog posts, course outlines, conversation logs, wiki chunks.

Read the material. Identify skill candidates based on pattern recognition from existing skills in this workspace. For each candidate:

### [Candidate Name]
- **What it does:** One sentence
- **Source:** Where in the material this was found
- **Hallmarks:** Which taxonomy items (1-6) it exhibits
- **Input → Output:** What goes in, what comes out
- **Frequency:** daily / weekly / per-project / one-time

Phase 2: Score and Rank

Leverage × Frequency × Promptability = Extraction Priority

High-priority (50+): build immediately. Medium (25-49): build when needed. Low (<25): note or discard.

Phase 3: Spec the Winners

For each high-priority candidate, draft a skill spec for the skill-creator:

## Skill Spec: [name]

**Purpose:** What this skill accomplishes
**Trigger:** What would the user say to invoke this?
**Not for:** What this skill should NOT be used for

**Input → Output**
**Workflow:** [numbered steps]
**Bundled Resources:** scripts, references, assets needed
**Tool Dependencies:** MCPs, APIs, CLI tools required
**Examples Needed:** What examples would make this work well?
**Source Attribution:** Where this was extracted from

Phase 4: Present to User

1. Top 5 high-priority skills with full specs
2. Medium-priority as one-liners
3. Non-skill material routed elsewhere

Ask: "Which should I build first?"

Routing Non-Skills

Test Corpora

Already chunked and indexed in wiki-projects:

• Kallaway — 66 chunks, framework-dense → should yield many skills
• Jenny Hoyos — 60 chunks, short-form methodology → should yield skills
• Colin & Samir — 119 chunks, strategy + interviews → mixed

Related Skills

• skill-creator — Takes a spec and builds the full skill. Extractor feeds into creator.
• voice-analyzer / voice-wizard — Specialized extractors for voice/style skills.
• anti-ai-writing — Example of a well-built skill.
• book-chapter-writer — Example of a complex multi-phase workflow skill.

Extract the workflow, not the knowledge. If Claude can DO it repeatedly with different inputs, it's a skill. If Claude can only KNOW it, it's a reference doc.