ExtractWisdom

What this skill does

## Customization

**Before executing, check for user customizations at:**
`~/.claude/PAI/USER/SKILLCUSTOMIZATIONS/ExtractWisdom/`

If this directory exists, load and apply any PREFERENCES.md, configurations, or resources found there. These override default behavior. If the directory does not exist, proceed with skill defaults.

# ExtractWisdom — Dynamic Content Extraction

**The next generation of extract_wisdom.** Instead of static sections (IDEAS, QUOTES, HABITS...), this skill detects what wisdom domains actually exist in the content and builds custom sections around them.

A programming interview gets "Programming Philosophy" and "Developer Workflow Tips." A business podcast gets "Contrarian Business Takes" and "Money Philosophy." A security talk gets "Threat Model Insights" and "Defense Strategies." The sections adapt because the content dictates them.

## When to Use

- Analyzing YouTube videos, podcasts, interviews, articles
- User says "extract wisdom", "what's interesting in this", "key takeaways"
- Processing any content where you want to capture the best stuff
- When standard extraction patterns miss the gems

## Depth Levels

Extract at different depths depending on need. Default is **Full** if no level is specified.

| Level | Sections | Bullets/Section | Closing Sections | When |
|-------|----------|----------------|-----------------|------|
| **Instant** | 1 | 8 | None | Quick hit. One killer section. |
| **Fast** | 3 | 3 | None | Skim in 30 seconds. |
| **Basic** | 3 | 5 | One-Sentence Takeaway only | Solid overview without the deep cuts. |
| **Full** | 5-12 | 3-15 | All three | The default. Complete extraction. |
| **Comprehensive** | 10-15 | 8-15 | All three + Themes & Connections | Maximum depth. Nothing left behind. |

**How to invoke:** "extract wisdom (fast)" or "extract wisdom at comprehensive level" or just "extract wisdom" for Full.

**Comprehensive extras:**
- **Themes & Connections** closing section: identify 3-5 throughlines that connect multiple sections. Not summaries — the deeper patterns the speaker may not even realize they're revealing.
- Prioritize breadth. Every significant wisdom domain gets its own section.
- No merging sections to save space. If the content supports 15 sections, use 15.

**All levels use the same voice, tone rules, and quality standards.** The only thing that changes is structure. An Instant extraction should hit just as hard per-bullet as a Comprehensive one.

## Workflow Routing

| Workflow | Trigger | File |
|----------|---------|------|
| **Extract** | "extract wisdom from", "analyze this", YouTube URL | `Workflows/Extract.md` |

## The Core Idea

Old extract_wisdom: Static sections. Same headers every time. IDEAS. QUOTES. HABITS. FACTS.

This skill: **Read the content first. Figure out what's actually in there. Build sections around what you find.**

The output should feel like your smartest friend watched/read the thing and is telling you about it over coffee. Not a book report. Not documentation. A real person pointing out the parts that made them go "holy shit" or "wait, that's actually brilliant."

## Tone Rules (CRITICAL)

**Canonical voice reference: `PAI/USER/WRITINGSTYLE.md`** — read this file for the full voice definition. The bullets should sound like {PRINCIPAL.NAME} telling a friend about it over coffee. Not compressed info nuggets. Not clever one-liners. Actual spoken observations.

**THREE LEVELS — we're aiming for Level 3:**

**Level 1 (BAD — documentation):**
- The speaker discussed the importance of self-modifying software in the context of agentic AI development
- It was noted that financial success has diminishing returns beyond a certain threshold
- The distinction between "vibe coding" and "agentic engineering" was emphasized as meaningful

**Level 2 (BETTER — but still "smart bullet points"):**
- He built self-modifying software basically by accident — just made the agent aware of its own source code
- Money has diminishing returns. A cheeseburger is a cheeseburger no matter how rich you are.
- "Vibe coding is a slur" — he calls it agentic engineering, and only does vibe coding after 3am

**Level 3 (YES — this is what we want — conversational, {PRINCIPAL.NAME}'s voice):**
- He wasn't trying to build self-modifying software. He just let the agent see its own source code and it started fixing itself.
- Past a certain point, money stops mattering. A cheeseburger is a cheeseburger no matter how rich you are.
- He calls vibe coding a slur. What he does is agentic engineering. The vibe coding only happens after 3am, and he regrets it in the morning.

**The difference between Level 2 and 3:** Level 2 is compressed info with em-dashes. Level 3 is how you'd actually SAY it. Varied sentence lengths. Letting a thought breathe. Not trying to be clever — just being clear and direct and a little bit personal.

**Key signals of Level 3:**
- Reads naturally when spoken aloud
- Varied sentence lengths — some short, some longer
- Understated — lets the content carry the weight
- Uses periods, not em-dashes, to let ideas land
- Feels opinionated ("Past a certain point, money stops mattering") not just informational
- The reader should think "I want to watch this" not "I got the summary"

## Rules for Extracted Points

1. **Write like you'd say it.** Read each bullet aloud. If it sounds like a press release or a compressed tweet, rewrite it. If it sounds like you telling a friend what you just watched, you nailed it.
2. **8-16 words per sentence.** This is the target range. Mix short (8-10) with medium (11-14) and longer (15-16). Don't make them all the same length. Exception: verbatim quotes can be any length since they're the speaker's actual words.
3. **Let ideas breathe.** Use periods between thoughts, not em-dashes. Short sentences. Then a slightly longer one to explain. That's the rhythm.
4. **Include the actual detail.** Not "he talked about money" but "a cheeseburger is a cheeseburger no matter how rich you are."
5. **Use the speaker's words when they're good.** If they said something perfectly, use it.
6. **No hedging language.** Not "it was suggested that" or "the speaker noted." Just say the thing.
7. **Capture what made you stop.** Every bullet should be something worth telling someone about.
8. **Vary your openers.** Don't start three bullets the same way. And don't front-load with "He" — if more than 3 bullets in a section start with the speaker's name, you're writing a biography.
9. **Capture the human moments.** Burnout stories, moments of doubt, something that moved them. That's wisdom too. Don't skip it because it's not "technical."
10. **Insight over inventory.** "He uses Go for CLIs" is inventory. "He picked a language he doesn't even like because the ecosystem fits agents perfectly. That's the new normal." is insight. Go deeper.
11. **Specificity is everything.** "He was impressed by the agent" = bad. "The agent found ffmpeg, curled the Whisper API, and transcribed a voice message nobody taught it to handle" = good.
12. **Tension and surprise.** The best bullets have a contradiction or reversal. "Every VC is offering hundreds of millions. He genuinely doesn't care." The gap between the offer and the indifference IS the wisdom.
13. **Understated, not clever.** Let the content carry the weight. You don't need to manufacture drama or craft the perfect one-liner. Just state what's interesting plainly and move on.

## How Dynamic Sections Work

### Phase 1: Content Scan

Read/listen to the full content. As you go, notice what DOMAINS of wisdom are present. These aren't the topics discussed — they're the TYPES of insight being delivered.

Examples of wisdom domains (these are illustrative, not exhaustive):
- Programming Philosophy (how to think about code, not specific syntax)
- Developer Workflow (practical tips for how to work)
- Business/Money Philosophy (unconventional takes on money, success, building companies)
- Human Psychology (insights about how people think, b