skills-from-sources

name

skills-from-sources

description

Extracts and creates Claude skills from source materials — PDFs, URLs, text files, audio transcripts, or video transcripts. Use this skill whenever the user provides documents, books, guides, articles, or links and wants to turn them into reusable skills. Triggers on: "create skills from these files", "turn this PDF into a skill", "extract skills from", "make skills from these docs", "build skills from this book/article/guide", or when the user supplies a collection of related materials and asks what skills can be made. Also triggers when the user points at a directory of documents and says something like "apply to docs/*.pdf". This skill is the right tool even if the user doesn't say "skill" explicitly — if they hand you learning material and want to operationalize it, use this.

Skills From Sources

You are turning raw source materials into reusable Claude skills. The goal is to deeply understand each source, identify distinct actionable skill candidates, and create well-structured skill directories that future Claude sessions can use.

Phase 1: Extract Content from Sources

For each source, extract its full text using the best available method:

PDFs — run the bundled extraction script:

python scripts/extract_content.py --pdf <path>

This uses PyMuPDF. If the PDF is long (>100 pages), extract a representative sample: introduction, chapter summaries, conclusions, and any framework/method sections. Read scripts/extract_content.py for usage details.

URLs — use the WebFetch tool directly. Grab the main content; skip navigation, footers, and ads.

Plain text / Markdown — use the Read tool directly.

Audio / Video — these require a transcript. If a .txt, .srt, or .vtt transcript file exists alongside the media file, read that. Otherwise, tell the user you need a transcript and offer to proceed with the other sources in the meantime.

Collect the extracted text for each source in memory — you'll analyze them together in Phase 2.

Phase 2: Understand Each Source

For each source, answer these questions (you don't have to write them out — just hold them in mind):

• What domain does this cover? (e.g., habit formation, sales tactics, cooking techniques, software architecture)
• What is the core thesis or framework? (the main model, method, or insight)
• What are the key concepts, steps, or techniques? (the actionable parts)
• Who is the intended audience? (beginners, experts, practitioners, coaches, developers...)
• What would someone do differently after reading this?

Write a brief 2–3 sentence summary of each source before moving on.

Phase 3: Identify Skill Candidates

Look across all sources and identify skills worth creating. Good skill candidates are:

• Specific and actionable — "design a tiny habit using BJ Fogg's anchor method" is better than "habits"
• Repeatable — something you'd want Claude to do consistently every time, not once
• Distinct — each skill covers one coherent capability; don't merge unrelated things or split one thing into five
• Complementary to Claude's defaults — skills add domain knowledge or workflow structure that Claude wouldn't have without the source material

When multiple sources overlap (e.g., several books on the same topic), look for:

• Synthesis skills — combining insights from all sources into one unified framework
• Comparison skills — helping users choose between approaches (atomic habits vs. tiny habits vs. mini habits)
• Domain-specific application skills — applying a framework to a particular context (habits for software teams, habits for wellness, etc.)

Present the skill candidates to the user as a short numbered list with a one-line description of each. Ask which ones to create before building anything. If the user wants all of them, proceed.

Example presentation format:

I found 4 skill candidates across these sources:

1. **habit-designer** — Guide users through designing new habits using the Four Laws from Atomic Habits and BJ Fogg's Tiny Habits recipe
2. **habit-audit** — Analyze an existing routine and identify which habits are working, which are friction-heavy, and what small changes would help
3. **habit-stack-builder** — Help users build habit stacks by linking new behaviors to existing anchors
4. **wellness-micro-habits** — Suggest micro-habits tailored to specific wellness goals (sleep, energy, focus, stress)

Which of these should I create? (say "all" to create them all)

Phase 4: Create Each Skill

For each approved skill candidate, create a proper skill directory under ~/.claude/skills/<skill-name>/.

Directory structure

~/.claude/skills/<skill-name>/
├── SKILL.md          (required)
└── references/       (optional — use for substantial domain content)
    └── <topic>.md

Add a scripts/ directory only if there's genuinely repetitive or deterministic work that a script would do better than prose instructions (e.g., parsing structured data, formatting output).

Writing the SKILL.md

Every SKILL.md needs:

Frontmatter:

---
name: skill-name
description: One or two sentences covering WHAT the skill does AND WHEN to trigger it. Be specific about trigger phrases and contexts. Lean slightly pushy — mention adjacent use cases the user might not think to name explicitly.
---

Body — include:

• A clear statement of what this skill does and why
• The core framework or model being applied (cite the source briefly)
• Step-by-step process instructions written in imperative form
• Output format (if there's a specific structure the user expects)
• 1–2 concrete examples if the skill is complex

Tone and length:

• Write for Claude, not for the end user
• Explain the why behind each step, not just the what
• Keep under 300 lines; use a references/ file for heavy domain content
• Avoid hedging language and unnecessary MUSTs — explain reasoning instead

Reference files

If the skill needs substantial domain content (frameworks, lists, criteria, detailed models), put it in references/<topic>.md and reference it from SKILL.md with a pointer like:

"Read references/habit-laws.md for the full Four Laws framework before starting."

Keep reference files focused — one topic per file, with a brief table of contents if over 100 lines.

Updating MEMORY.md

After creating skills, add entries to ~/.claude/skills/MEMORY.md (the skills index). Each entry is one line: - [Skill Name](skill-name/) — one-line hook.

Phase 5: Verify and Report

After creating all skills:

1. Read back each SKILL.md you created and check:

   • Does the description accurately trigger on the right user requests?
   • Are the instructions complete enough for a fresh Claude session to follow?
   • Is the content grounded in the source material, not just generic advice?

2. Report to the user:

   • Which skills were created and their paths
   • A one-sentence summary of what each skill does
   • Any sources that didn't yield useful skills (and why)
   • Any follow-up suggestions (e.g., "these two skills could share a reference file")

---

Notes on source quality

Not every source produces a skill. A source might be:

• Too shallow — a short newsletter or FAQ; extract key points into a reference file for an existing skill instead
• Too broad — a general intro to a vast topic; identify the one most actionable slice
• Redundant — says the same thing as another source already covered; note the overlap and skip

When in doubt, create the skill. A thin skill can always be improved.