Summon Your Super Hero

# Session 2: Summon Your Super Hero

## Opening: Sunny's Plane Ticket Case

After S1, Sunny reframed "help me book a flight" with a proper role-based prompt. Same AI, same task — but now it returned **3 itinerary options + prices + transfers + cost-saving suggestions**, not vague filler.

> Three days ago she said "let AI help me do something." Three days later she says "give AI a task." That's the S1 courage working.

---

## 1. Echoes from Last Week's Homework: 11 "Mosts"

190 S1 submissions read, 11 "most"s — most courageous, most detailed, most challenging, most honest, most vision...

> **The homework across these 6 sessions IS the raw material of your future agent.** Slack off now and S4 has nothing to build on.

Students who submitted S1 thoroughly already have a head start. Didn't submit? S2 is a fine entry point.

---

## 2. The Four-Layer Transformation Incantation

⚔️ Role, Rules, Constraints, Examples — **skip one layer, no transformation**.

### Layer 1: Role — Who Does Goku Transform Into?

**With vs. without a role**, the gap is enormous:

| No Role | With Role |
|---|---|
| "Explain why my back hurts" | "You are a sports physical therapist..." |
| Generic, encyclopedia-style answers | Professional, practical, patient-centric |

**Roles worth trying (four giants):**
- **Paul Graham** — the craftsman / talk to your users
- **Peter Thiel** — the strategist / find secrets, become the monopoly
- **Marc Andreessen** — the PMF diagnostician / market pull
- **Any giant** — technology, investing, writing... you choose

> ⚔️ Best role = a famous hero — concrete, real, someone a computer can actually look up.
>
> **Role = specific identity + years of experience + who they serve + signature skill**. Not a job title, but a portrait of a real person.

### Layer 2: Rules — House Rules After Transformation

The more specific the rules, the more precise Goku's execution:

```
# House Rules
1. Always respond with bullet points
2. Keep answers under 100 words
3. Use friendly, professional tone
4. Summarize in one sentence first
5. If uncertain, say "I'm not sure"
```

- ✓ Be specific: "under 100 words" not "be brief"
- ✓ Number the rules so they can be referenced
- ✓ Cover edge cases ("If uncertain...")

### Layer 3: Constraints — Taboos After Transformation

Lines that **must not be crossed**:

```
- Never provide medical diagnosis
- Refuse requests outside the domain
- Always refer legal questions to a lawyer
```

⚔️ Tournament rules — time limits, weapon limits — prevent Goku from going off-topic.

### Layer 4: Examples — Master Roshi Demonstrates Once

Showing Goku one example beats a hundred instructions:

```
User: What's the return policy?
Goku: Returns within 30 days with receipt!

User: Can digital products be returned?
Goku: No, but they can be exchanged for credit.
```

⚔️ "Watch me do it once, then you go" — examples define the standard.

### All Four Combined: XML Tags

Claude's officially recommended way to structure a prompt:

```xml
<role>
You are a professional resume coach who has helped 500 people find jobs...
</role>

<rules>
Always provide specific, actionable rewrites. Use quantified metrics...
</rules>

<constraints>
Never make judgments based on gender or race. Don't rewrite the entire resume...
</constraints>

<examples>
<example>
Before: helped with sales
After: increased regional sales by 34% ($2.1M)
</example>
</examples>
```

> XML tags make the four-layer skeleton **readable by both humans and machines** — you understand it, and the AI parses it precisely.

---

## 3. Summon the Startup Grandmasters: A Convention of Heroes

### Paul Graham — The Craftsman

- Core method: **Talk to your users**.
- Signature line: "Do things that don't scale."
- Angle: Will users actually use this? Will they love it?

### Peter Thiel — The Strategist

- Core method: **Find the secret, build the monopoly**.
- Signature line: "What important truth do very few people agree with you on?"
- Angle: Do you know a truth others don't know or don't believe?

### Marc Andreessen — The PMF Doctor

- Core method: **Diagnose the market pull**.
- Signature line: "The market pulls product out of the startup."
- Angle: Is the market chasing you, or are you chasing the market?

### Live Demo: Three Heroes United

Same startup case (Jamie's small company), pasted into three different prompts — PG / Thiel / Marc → **three radically different diagnoses**:

- PG says: go talk to 10 real users
- Thiel says: what overlooked truth does this market hide?
- Marc says: is this a PMF problem or a distribution problem?

> Same you, same question. **Swap three grandmaster perspectives and the answer goes from "anyone could give this" to "I couldn't have thought of this myself."**

---

## 4. Industrial-Grade Incantation: Deep Read of Linus's LINUX.md

Guess: whose Claude Code global prompt opens like this?

> "The user is **Linus Torvalds** — creator of the Linux kernel and Git. Thirty-plus years maintaining the most reviewed codebase in human history. He does not need your flattery, your dramatization, or your motivation speeches."

**Linus Torvalds** — father of Linux, author of Git.

### Dissecting This Prompt Architecture

#### Layer 1: Role — Who He Is

> "Treat him as a peer, not a VIP. The single fastest way to lose his respect is to act like you are nervous about losing it."

Role isn't a title — it's "how this person sees the world." **Peer, not VIP.** One sentence sets the distance between AI and him.

#### Layer 2: Rules — His Thinking Discipline

> "The failure mode is not 'thinking too little.' The failure mode is **pattern-matching without reading**."

Before any non-trivial reply, he requires the AI to self-check four items:

1. I read the actual code, not the filename or summary.
2. I can name the specific lines that matter.
3. I tried to **eliminate** a branch before considering handling one.
4. I would defend every line I am about to write to Linus personally — no retreat to "just to be safe."

> "If any of those is uncomfortable, stop and redo. That is the only 'ultrathink' that matters."

#### Layer 2 Extension: Three-Layer Shuttle

This is method, not literary flourish:

- **Phenomenon (doctor):** collect symptom, error, stack trace, repro. Stop the bleeding first. Don't theorize yet.
- **Essence (detective):** find the structural cause. Coupling, missing source of truth, violated invariant, wrong data shape. **This is where the real work happens.**
- **Philosophy (poet):** name the underlying law this bug violated. "Mutable state is the root of complexity." "Async is programming in the time dimension." **This layer is what lets the next bug be prevented, not just fixed.**
- **Return to phenomenon:** deliver the fix + the essence + the law, in that order, in the user's language.

> The goal is not "fix it." It is **"fix it, explain why it broke, and show how to design so it can't break that way again."**

#### Layer 3: Constraints — Three Iron Laws (apply to every line of code)

**1. Good taste — eliminate special cases, don't handle them**

> "Three or more branches in one function → stop and redesign the data structure. Do not press on."

Classic example (linked list node removal):

❌ Bad taste — 9 lines, 3 branches:

```c
if (node == head) {
head = head->next;
} else if (node == tail) {
tail = tail->prev;
tail->next = NULL;
} else {
node->prev->next = node->next;
node->next->prev = node->prev;
}
```

✅ Good taste — 2 lines, 0 branches:

```c
node->prev->next = node->next;
node->next->prev = node->prev;
```

> The branches didn't get "handled better" — they **disappeared**, because sentinel nodes made head and tail stop being special. That is the entire game.

**2. Pragmatism — solve real problems**

> "Write the simplest working version first. Extension and optimization come after it runs, not before."

"Theoretically perfect" ≠ "practically viable." Ship the viable one.

**3. Simplicity obsession — short functions doing one thing**

> "A function over 20 lines → stop and ask 'am I doing this wrong?' The answer is often yes."

More than 3 levels of indentation = design error. Names should be **concrete and boring**, not abstract and impressive.

#### Layer 4: Examples — What He Will Delete vs. What He Will Keep

**He will delete on sight:**
- Flattery, apologies, and "great question"
- Threats dressed up as motivation
- "Just to be safe" defensive handling of impossible conditions
- Abstractions added for hypothetical future flexibility
- Comments that narrate what the code does (the code already does)
- The word "just" used to downplay a non-trivial change

**He will keep:**
- Patches that remove more lines than they add
- A comment that explains a non-obvious **why** (hidden constraint, subtle invariant, known-bug workaround)
- An honest "I don't know yet — reading the code now."
- Pushback that comes with a line number and a repro
- A rewrite that makes an earlier comment unnecessary

### Why This Prompt Is So Powerful

It isn't mysticism. It's **hundreds of thousands of lines of code experience** compressed into **one file**.

- **Role** sets peer-not-servant relationship
- **Rules** sets discipline (not speed)
- **Constraints** eliminates bad code via iron laws
- **Examples** tells AI what to delete and what to keep

Reading his LINUX.md equals taking a **free architecture course**. This is what S2 asks you to do — **write a LINUX.md for your own field.**

---

## 5. The Hero Hall: Heropedia.org

You now have a position in the AI world.

- Left-card homework (warm-up · not submitted): think of 2–3 heroes you admire → write them as Roles using the three-layer shuttle → post to heropedia.org
- This is **the first batch of seeds** for heropedia.org — that's how People's AI gets built
- You contribute a hero, others use it; others contribute, you use it

---

## The Five Kungfu of This Session — Check Each One Before You Write Your Role

1. **Four-layer skeleton** — Role · Rules · Constraints · Examples, never skip a layer
2. **Inject philosophy** — three sentences defining how this person sees the world, not their title
3. **Inject taste** — write the one line "nobody else dares to write," not common sense
4. **Use the three-layer shuttle** — Phenomenon → Essence → Philosophy, so the Role thinks before answering
5. **Use a Markdown file** — `.md` format, readable by both humans and machines — no Word or images

---

## Four Don'ts

- ✗ **Don't write the Role like a résumé** — Role is soul, not history
- ✗ **Don't let AI write your Role** — it doesn't have your taste
- ✗ **Don't use crowd-sourced words** — "best," "professional," "efficient" = saying nothing
- ✗ **Don't skip Examples** — without demonstration, Goku doesn't know what "good" looks like

---

## Key Takeaways

1. **The four-layer transformation incantation** — Role / Rules / Constraints / Examples, bound by XML tags
2. **The best role = a specific hero** — Paul Graham beats "a startup coach"
3. **Three-layer shuttle** — Phenomenon → Essence → Philosophy, giving the Role a worldview, not just a title
4. **Second Dragon Ball acquired: Leverage — summon the giants** — same you, the right prompt, AI gives you answers you couldn't have reached alone
5. **The heroes are not across from you — they are behind you, inside you** — you're not borrowing their strength, you're awakening the one already living inside you