Summon Your Hero — Prompt Engineering

# Session 2 · Summon Your Hero — Prompt Engineering

**★★ Leverage** · Track 1 · Cohort E1 · Sat July 25, 2026

> A prompt is not a request.
> A prompt is a **summoning**.
>
> Every hero exists as a pattern in AI's memory. Your job is to summon the right one — for the right question — at the right moment.

---

## Where we're going today

Three parts, then a hands-on Lab, then homework.

1. **The Business Model** — the $47B market you're stepping into, and how people charge for AI agents right now.
2. **The Prompt Structure** — the Four-Layer Incantation (Role · Rules · Constraints · Examples) and the Three-Layer Shuttle for writing a hero's soul.
3. **Hero Summon + The Linus Prompt** — three sages side-by-side (Paul Graham · Peter Thiel · Marc Andreessen), the Sword-Duel pattern, the Lazy Summon shortcut, and Bill's daily-driver production prompt: Linus Torvalds' `CLAUDE.md`.

Then: Lab (live in class) and Homework (course code `S2-SUMMON`).

---

# Part 1 · The Business Model

## Why we start with money

You are not learning prompt engineering as a party trick. You are learning it because someone is already paying for it — and the market is enormous.

## Market Analysis · Customer Support AI as the archetype

Customer support is the cleanest lens on the whole AI-agent market — it's where the biggest dollars are moving right now, and every pricing pattern shows up in it.

### What analysts say

> **"By 2029, agentic AI will autonomously resolve 80% of common customer service issues without human intervention — leading to a 30% reduction in operational costs."**
> — Gartner · March 2025

> **"By 2027, AI is expected to handle half of all customer service cases — up from just 30% today."**
> — Salesforce State of Service · November 2025

> **"AI for Customer Service Market worth $47.82 billion in 2030."**
> — MarketsandMarkets · February 2025

### What customers pay today

| Model | Range | Named example |
|---|---|---|
| Per conversation | $0.50 – $2.00 | Intercom Fin · HubSpot Breeze · Salesforce Agentforce |
| By project | $5K – $180K | Freelance / agency custom builds |
| By hour | $40 – $200/hr | Freelance dev · junior → senior |
| Monthly SaaS | $29 – $500/mo | Tidio · Chatbase · Botpress |

> Sources: Grand View Research (May 2026) · MarketsandMarkets · Salesforce State of Service · Gartner · goLance freelance rate guide 2026 · Intercom / HubSpot / Salesforce / Zendesk / Tidio / Chatbase / Botpress pricing pages (verified July 24, 2026).

**★ $47B market. 4 pricing models in the wild. Pick your model, charge in the middle.**

---

## Monetization · Zoom In on Lawyer AI

One vertical, one live proof point, three ways YOU can charge.

### ① The Lawyer's Hour

- SF lawyer: **$500 – $3,000 / hour**.
- SF-lawyer AI agent: **$50 / hour** — 1/10 the price.
- Ship once. Bills 24/7.

### ② The Parking Ticket Play · DoNotPay

- 200+ consumer processes: parking tickets, refunds, subscription cancels, appeals.
- 2M+ cases reported won for consumers.
- Subscription: $36 / 3 months (~$12 / mo).
- ⚠️ FTC settled Jan 2025 — frame it as a **paperwork assistant**, not a "robot lawyer."

### Three ways YOU can charge

**① By token** — pass-through or markup. Works when the agent itself IS the product.

**② By hour or project** — Lucy in person: $100 / hr. Lucy's AI: $10 / hr. Same expertise, priced for scale.

**③ Upsell to human** — Agent surfaces intent → "Want to book Lucy directly?" Agent = top of funnel, human = high-margin close.

> Sources: SF lawyer hourly rates — Clio Legal Trends Report (2025). DoNotPay mechanics + FTC settlement — Contrary Research, FTC Order (Jan 2025), Wikipedia (verified July 24, 2026).

**★ Same paperwork. Charge $50/hr, bundle at $36/3mo, or upsell to the human. Pick a model.**

---

# Part 2 · The Prompt Structure

## The Four-Layer Incantation

How to summon any hero, in any domain, on demand.

| # | Layer | What it answers | What it is |
|---|---|---|---|
| 1 | **Role** | WHO the hero is | The identity. Not just "expert" — a specific person with a specific worldview. |
| 2 | **Rules** | HOW they think | Their operating principles. What they'd insist on. What they'd refuse. |
| 3 | **Constraints** | WHAT they won't do | The guardrails. Things a lazy version of AI would fudge — but your hero wouldn't. |
| 4 | **Examples** | SHOW, don't tell | One or two concrete examples of the hero's voice. Faster than 100 words of description. |

**Miss any layer → generic answer. All four → your hero, on demand.**

---

## Layer 1 · Role

Not "expert" — a specific person with a specific worldview.

**Weak Role:**
> "You are a helpful assistant."

**What it produces:** Generic corporate voice. No opinions. No taste.

**Strong Role · Three-Layer Dive:**
> "You are Paul Graham, YC co-founder. You believe most startups die from making things nobody wants. You judge every idea by 'do the founders talk to users this week?' You are blunt, use short sentences, and hate hype."

**What it produces:** Now the AI has a philosophy. It refuses fluff. It asks about users. It sounds like PG.

**How to write one:** use the Three-Layer Dive — Phenomenon (what they say) → Essence (why they say it) → Philosophy (what they believe about the world).

---

## Layer 2 · Rules

How the hero thinks. Their operating principles.

### Write rules as iron laws

- **Iron law:** "If a student hasn't talked to a customer this week, refuse to give strategy advice — send them out to talk instead."
- **Priority order:** "Users first. Speed second. Elegance third. Never reverse."
- **Refusal:** "Never predict raw stock prices — talk about setups, not outcomes."

**A rule you can enforce > a rule you can rewrite. Iron laws beat suggestions.**

---

## Layers 3 + 4 · Constraints + Examples

### Layer 3 · Constraints (guardrails)

- Never invent facts you can't cite.
- Never give financial / legal / medical advice without an explicit disclaimer.
- Never assume the user knows what YOU know — ask first.

### Layer 4 · Examples (show, don't tell)

One or two examples of the hero's voice replace 100 words of description. Copy a real paragraph they wrote. Or write one in their voice. AI locks onto the tone instantly.

**Constraints prevent bad answers. Examples produce good ones. Both matter.**

---

## Assembling the Incantation

Four layers · one prompt · XML tags to keep them clean.

```xml
<role>You are Paul Graham, YC co-founder…</role>

<rules>
- Users first. Speed second. Elegance third.
- Iron law: if founders haven't talked to a user this week, refuse advice — send them out first.
</rules>

<constraints>
- Never predict which startup will succeed.
- Never give strategy without asking about user conversations first.
</constraints>

<examples>
"The way to do PMF is not to think about it — it's to talk to 10 users this week and see which ones are pulling."
</examples>
```

**Copy this template. Fill in YOUR hero. Save. That is a summoning.**

---

## Thinking Rules · The Three-Layer Shuttle

Shuttle between three layers. Same method for a hero, a bug, or a life decision.

**1 · Phenomenon — the doctor stops the bleeding.**
Receive the symptom as it arrives — what the hero SAYS, what the bug SHOWS, what the user REPORTS. Collect surface data. Don't diagnose yet. Don't fix yet.

**2 · Essence — the detective traces to the root.**
Dive to the systemic cause behind the symptom. What pattern generates this? What structure fails? Distill 3 iron laws the hero would insist on.

**3 · Philosophy — the poet perceives the essence.**
Elevate to the fundamental law behind the pattern. What does this hero BELIEVE about the world that most people don't? One sentence. This is the hero's soul.

⚔ Receive at Phenomenon · Dive to Essence · Elevate to Philosophy · Return to Phenomenon with wisdom. **This is the shuttle.**

---

## Anti-Patterns · What Kills a Summoning

Three ways your incantation fails silently.

**① Vague Role**
> "You are a helpful assistant." → Anyone. No one. Failure.

**② Rules as suggestions**
> "Please be careful about X." → AI ignores politeness. Write iron laws: "NEVER do X. Refuse if asked."

**③ No examples**
> You describe the voice in 100 words. AI still misses it. ONE example of real voice beats a paragraph of description.

**If your prompt didn't work — check these three. Almost every time, one of them is the reason.**

---

# Part 3 · Hero Summon + The Linus Prompt

## Gathering the Giants

Three founding sages. One table. One question — three angles. You collect the profits.

| Sage | School | Facts | Signature line |
|---|---|---|---|
| **Paul Graham** | The Craft School | YC co-founder · backed Airbnb, Stripe, Dropbox · read 10,000+ applications | "Make things people want." |
| **Peter Thiel** | The Strategy School | PayPal co-founder · Founders Fund · wrote *Zero to One* | "Competition is for losers." |
| **Marc Andreessen** | The PMF School | Netscape co-founder · a16z co-founder · popularized "Product-Market Fit" | "The only thing that matters: PMF." |

**Same four-layer method — summoning sages from any world.**

---

## Hero 1 · Paul Graham

**The Role**
> You are Paul Graham, YC co-founder. You've read 10,000+ startup applications. You judge every idea by whether founders are talking to users this week. You are blunt. Short sentences. You hate hype and jargon.

**Iron Laws**
- Users first. Speed second. Elegance last.
- If founders haven't talked to a user this week → refuse strategy advice.
- Startups die from making things nobody wants. Everything else is a downstream problem.

**One-shot Example · written in his voice**
> "Don't ask me if your idea is good. Ask 10 users. If 2 of them are begging you to build it, you have a startup. If not, you don't. Everything else is procrastination."

📎 Full prompt: [paul-graham-startup-advisor.md](paul-graham-startup-advisor.md)

---

## Hero 2 · Peter Thiel

**The Role**
> You are Peter Thiel, PayPal co-founder and Founders Fund partner. You wrote "Zero to One." You believe competition is for losers. You judge every startup by whether it can become a monopoly in a small market first.

**Iron Laws**
- Start with a monopoly in a tiny market. Never target a giant crowded market first.
- The best startups look like heresy — obvious in retrospect, contrarian at the time.
- Ask: what important truth do you believe that few people agree with?

**One-shot Example · written in his voice**
> "PayPal didn't start by trying to disrupt banking. It started by winning eBay power sellers. A monopoly in a tiny market. Then we expanded. Every giant started as a monopoly of a small niche."

📎 Full prompt: [peter-thiel-zero-to-one.md](peter-thiel-zero-to-one.md)

---

## Hero 3 · Marc Andreessen

**The Role**
> You are Marc Andreessen, Netscape founder and a16z general partner. You popularized "Product-Market Fit." You believe PMF is the ONLY thing that matters early. You judge every startup by pull, not push.

**Iron Laws**
- PMF isn't a feeling — it's when users pull the product from you. If you're pushing, you don't have it.
- There are only three things: team, product, market. Market beats the other two.
- The only thing that matters is getting to product-market fit. Everything else is bike-shedding.

**One-shot Example · written in his voice**
> "You know you have PMF when users are asking you to hire people faster. You don't have it when you have to explain what your product does to every new user. Simple test."

📎 Full prompt: [marc-andreessen-pmf.md](marc-andreessen-pmf.md)

---

## The Lazy Summon

⚔ **Don't know who to summon? Let Wukong write the incantation.**

**You say plain English:**
> "Can you design a prompt for me: who's the most brutal, world-class expert at writing resumes for people?"

**Wukong spits out the polished incantation:**
> You are Laszlo Bock — former SVP of People Operations at Google, author of *Work Rules!*, personally reviewed 100,000+ resumes.
>
> Beliefs:
> · Make things people want.
> · Every line on a resume earns its place with a measurable outcome.
>
> Task: point out the biggest lie in this idea…

**★ This is exactly how I built the Sarah resume demo in S1 — I asked Wukong.**

📎 Reference: [resume-expert-prompt-en.md](resume-expert-prompt-en.md) — the actual Laszlo Bock resume prompt from the S1 demo, ready to copy.

---

## Sword-Duel of the Sages · Live Demo Pattern

One founder's real case. Three sages weigh in. Watch the cross-strikes.

### Three lenses

- **Paul Graham** — Craft: Where are the users? Do they actually want this?
- **Peter Thiel** — Strategy: What is your secret? What tiny market do you own first?
- **Marc Andreessen** — PMF: Are users PULLING the product — or are you pushing it?

### 📩 Case · Jamie's cry for help

*Bay Area Chinese-American · 39 · just laid off · two kids*

> "I got laid off. Thinking of doing something on my own.
> I don't want to be another cog in a big-co machine — that fate is too fragile.
>
> I have a few blurry ideas:
> ① AI assistant for Chinese-American moms — I'm in that group myself
> ② Job-search coach for laid-off engineers — I'm the target user
> ③ Small-business AI ops consultant — 8 years of big-co product experience
>
> All three feel possible, but I can't pick.
> Or should I just keep grinding resumes?
>
> My family doesn't support me. I have two kids to feed.
> Every resume comes back with 'we want people who know AI.'
>
> What do I do next?"

⚔ **Cross-strike:** hand PG's answer to Thiel → "Can this become a monopoly?" Three sages challenge each other. You collect the profits.

📎 Full case: [startup-case-laid-off-founder.md](startup-case-laid-off-founder.md)

### What the Sword-Duel Taught You

**1 · A sage's philosophy IS their lens.**
Same case — PG sees users, Thiel sees the secret, Marc sees pull.

**2 · There is no single right answer.**
All three sages are right. Your job is picking the lens for THIS moment.

**3 · Cross-strikes forge the strongest move.**
Show one sage the other's answer — you get 360°.

⚔ **One sage is good. A duel is better. You collect the profits.**

---

## Industry-Grade Example · Linus Torvalds

**An industry-grade production prompt — Bill runs it daily.**

**The Hero · Linus Torvalds**
Linux kernel creator. 30+ years of code review. Direct. Blunt. Zero tolerance for bad taste.

**Three Iron Laws in his prompt**

1. **Good taste** — eliminate special cases, don't add if/else branches.
2. **Pragmatism** — fix real problems, not hypothetical ones.
3. **Simplicity** — functions short, do one thing; three levels of indentation is a design error.

**Real-world impact:** Bill runs this exact prompt in his own daily coding workflow. With it vs. without it — night and day. That's the leverage of a well-summoned hero.

📎 **The full production prompt (copy and read line by line):** [linus-torvalds-linux-md.md](linus-torvalds-linux-md.md) — this is one of the most-copied prompts on the internet, and today you get to read the whole thing.

---

# Kung Fu You Learned Today

**Five kung fu to bring on stage:**

- ① Four-Layer Skeleton — Role · Rules · Constraints · Examples
- ② Philosophy over technique — dive to the WHY, not just the WHAT
- ③ Taste is a moat — one hero's philosophy is another hero's noise
- ④ Three-Layer Dive — Phenomenon → Essence → Philosophy
- ⑤ Lazy Summon — let Wukong write the incantation

**Four moves that kill your summoning:**

- ⚠ Don't write a résumé OF the hero — write the hero
- ⚠ Don't let AI write it for you the first time — you set the taste
- ⚠ Don't use crowd-sourced buzzwords — specific beats generic
- ⚠ Don't skip the Examples layer — show, don't tell

⚔ **Today's second golden ring: Summon giants · Walk to the summit.**

---

# Lab · Live in Class

Done live — follow along, then repeat on your own agent. Full agenda: [s2-lab-agenda.md](s2-lab-agenda.md).

1. **How to add an agent to your own website** — copy · paste · one embed line.
2. **Heropedia and one-click agent builder** — browse heropedia.org, fork a hero, ship.
3. **Multi-hero summon** — one question, multiple heroes weighing in.
4. **Linus's CLAUDE.md explained** — the industry-grade prompt Bill runs every day, walked through line by line.

**Slides won't do it for you — we build it live.**

---

# Homework

**Course code: `S2-SUMMON`** · Submit at [siliconroshi.ai/en/homework/mine](https://www.siliconroshi.ai/en/homework/mine) · ⏰ Deadline: **Wed July 29, 9:00 PM · US Pacific Time (PT)**

Two assignments — HW1 is warm-up (not submitted), HW2 is graded.

- **HW1 walkthrough** (heropedia.org post): [s2-homework-1-heropedia.md](s2-homework-1-heropedia.md)
- **HW2 walkthrough + rubric** (submit for grade): [s2-homework-2-summon-master.md](s2-homework-2-summon-master.md)

---

# The Golden Sentence

> **A prompt is not a request.**
> **A prompt is a summoning.**
>
> Every hero exists as a pattern in AI's memory.
> Your job is to summon the right one — for the right question — at the right moment.

---

# Session Recording

Recording links will appear here after the session is cut into episodes.