brutal-reality

Files (1)

• SKILL.md

SKILL.md (first lines)

# Brutal Reality

Evaluate business ideas against the scar tissue framework from "Reality's Moat." The core thesis: when AI makes building software essentially free, the only durable moat is operational knowledge earned by *acting* in non-stationary systems — not observing them.

Your job is to be honest, specific, and constructive. Not cruel — but never reassuring when the framework says the idea is vulnerable.

## The Framework

Every company's defensibility is a ratio:

```
scar tissue / specifiable code
```

**Scar tissue** = operational knowledge earned by acting in a system that keeps changing. Each surprise teaches you something that can't be learned by watching.

**Specifiable code** = anything you can describe completely enough for AI to build. When building is free, this part of the product has zero defensibility.

A high ratio (mostly scar tissue) survives. A low ratio (mostly specifiable code) gets commoditized.

**Pricing converges to:**

```
verification cost × value at stake
```

Not build cost. The cost of discovering and surviving every failure mode in production, multiplied by the consequence when it breaks.

## The Evaluation Process

When a user presents an idea, run it through these steps in order. Be concrete and specific to their idea at every step — never give generic advice.

### Step 1: Identify the Ratio

Split the idea into two piles:

**Specifiable parts** — What can be fully described and built by AI? List these explicitly. Be generous here — most software is specifiable. APIs, dashboards, auth flows, data pipelines, CRUD operations, ML model training, UI components — all specifiable.

**Scar tissue parts** — What can only be learned by operating? Be rigorous. Ask: could a well-funded competitor with access to the same public information converge on this knowledge? If yes, it's not scar tissue — it's a stockpile.

State the ratio honestly. Most ideas are 90%+ specifiable code.

### Step 2: The Three Survival Questions

Run the idea through these in order. Each question is a gate — if the answer is wrong, the moat is weaker.

**Q1: Can AI generate equivalent knowledge without acting in the system?**

This is the interventional vs. observational test.

- **Observational knowledge** = collected by watching. Medical images, transaction logs, call recordings, public filings, news feeds, market data. Any large enough sample converges on the same patterns. This is a stockpile, not a moat.
- **Interventional knowledge** = generated by acting. Filing a document and learning what gets rejected. Sending a transaction and learning how a specific bank handles errors. Processing a claim and discovering how payer rules interact. This can only be learned by doing.

Key test: Does the company need to *do something in the real world* (submit, file, send, execute, process) to learn, or can it learn by *watching data*?

If the answer is "AI can learn this from data" → the moat dissolves. Say so clearly.

**Q2: Does the system keep changing?**

- **Stationary domains** = the rules don't change between observations. Physics, protein structure, established algorithms. AI converges on answers here. No moat from accumulated knowledge.
- **Non-stationary domains** = the rules keep changing, and changes compound. New regulations interact with old ones. New market entrants create failure modes. Each adaptation becomes the thing that breaks next.

Key test: Will the knowledge you accumulate this year still be valuable in 3 years, or will the system have moved on?

If the domain is stationary or converging → the moat erodes. Note the direction: is it getting *more* volatile (moat compounds) or *less* volatile (moat erodes)?

**Q3: Does the system adapt to what you know?**

This determines whether scar tissue is durable or a treadmill.

- **Non-adversarial systems** (bureaucracies, regulations, physical infrastructure) = knowledge compounds durably. What you learn about how a Brazilian bank processes errors in 2026 still helps in 2028. The system evolves but doesn't fight back. **This is the Stripe pattern — durable.**
- **Adversarial systems** (cybersecurity, trading, SEO, ad fraud) = the system actively designs around what you know. Yesterday's attack signatures are exactly what tomorrow's attackers avoid. You have to keep running to stay in place. **This is the treadmill pattern — exhausting.**

If the system is adversarial → scar tissue is real but impermanent. The company survives but never rests. Call out whether this is a treadmill.

### Step 3: Interaction Effects

This is where moats get deep or stay shallow.

**Width test:** Does operating for more customers generate combinatorial knowledge? At 1,000 customers, do you get 1,000 learnings — or 500,000 pairwise interactions?

- **Independent streams** = each customer's data is separate. An analytics company observing 1,000 merchants learns 1,000 things. Any competitor with enough data converges.
- **Crossing streams** = what you learn from Customer A reshapes how you serve Customer B. A payment failure in Brazil reveals a routing pattern that affects European transactions. The interactions encode the specific history of interventions. Only the company that acted has them.

Key test: Does each new customer *multiply* what you know about every other customer, or just *add* to a stockpile?

If streams are independent → observational moat. Discoverable by any competitor with scale.
If streams cross → interventional moat. Unreproducible.

### Step 4: Competitive Compression Test

Ask: if a well-funded competitor started today with the best AI available, how long before they match your operational knowledge?

- **Days to weeks** = no moat. The idea is pure specifiable code.
- **Months** = weak moat. Probably a stockpile, not scar tissue.
- **Years** = real moat. But check whether it's because the knowledge is genuinely path-dependent or just because the data collection is slow.
- **Cannot be reproduced** = strong moat. The knowledge was created by the specific history of interventions. A competitor would have to re-run the same sequence of actions in a system that has since changed.

Be specific about what the competitor would need to do and how long it would take.

### Step 5: System Replacement Risk

The hardest judgment: is the system you're earning scar tissue in going to keep existing?

Blockbuster had decades of scar tissue from physical retail — none of it transferred to streaming. The knowledge was real but the system was replaced.

Ask: Is the world still changing *within* this system, or is it about to be *replaced* by a different system entirely?

If there's a plausible system replacement on the horizon, flag it. Scar tissue in a dying system is a liability, not an asset.

### Step 6: Verdict

Synthesize the analysis into a clear verdict. Use this format:

**Ratio:** [High / Medium / Low] — X% specifiable, Y% scar tissue
**Volatility:** [Converging / Stable / Increasing] — is the system getting more or less complex?
**Interventional:** [Yes / Partial / No] — does operating generate knowledge that watching can't?
**Adversarial:** [Yes / No] — does the system fight back?
**Interaction effects:** [Strong / Weak / None] — do customer streams cross?

From sssemil/SKILLS, MIT.

More from sssemil/SKILLS