Chapter 5: Mostly Harmless

“The Guide’s entry on Earth was originally one word: ‘Harmless.’ After extensive research, Ford Prefect updated it to ‘Mostly Harmless.’ This was considered a major upgrade.”

The same rating system applies to AI-generated code.

---

The Karpathy Scale

In February 2025, Andrej Karpathy coined the term “vibe coding”:

“There’s a new kind of coding I call ‘vibe coding’, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.”

He described a workflow where you:

• Talk to the AI with voice input
• “Accept All” without reading diffs
• Copy-paste error messages with no comment
• Build projects without really “coding”

This is liberating. This is powerful. This is also context-dependent.

Karpathy was building personal projects and prototypes. The stakes were low. The code was, in Guide terminology, Mostly Harmless.

But not all code is Mostly Harmless.

---

The Risk Classification System

This Guide proposes a four-tier classification system for vibe coding risk:

🟢 MOSTLY HARMLESS

Definition: If it breaks, nothing bad happens.

Examples:

• Personal side projects
• Prototypes and proofs of concept
• Learning exercises
• Internal tools with no data access
• Scripts you’ll run once

Vibe Level: Full Karpathy. Accept All. Embrace the vibes.

Verification: Run it, see if it works, move on.

🟡 CAUTION ADVISED

Definition: If it breaks, it’s annoying but recoverable.

Examples:

• Team internal tools
• Development environments
• Non-critical automation
• Content that will be reviewed before publishing
• Code that affects only you

Vibe Level: Moderate vibing. Skim the diffs. Test the happy path.

Verification: Quick review, basic testing, have a rollback plan.

🟠 DANGER

Definition: If it breaks, real problems occur.

Examples:

• Production code
• Code handling user data
• Financial calculations
• Authentication/authorization
• Anything with compliance requirements
• Public APIs

Vibe Level: Low vibing. Read every line. Question everything.

Verification: Code review, comprehensive testing, security audit, staged rollout.

🔴 HERE BE DRAGONS

Definition: If it breaks, catastrophic consequences.

Examples:

• Medical systems
• Financial trading
• Security infrastructure
• Safety-critical systems
• Anything where failure means lawsuits, injuries, or deaths

Vibe Level: No vibing. AI assists research only. Humans write and verify all code.

Verification: Multiple independent reviews, formal verification where possible, extensive testing, regulatory compliance.

---

The Tradeoff Matrix

As the Manning book Vibe Coding: Mistakes and Tradeoffs emphasizes, every vibe coding decision involves tradeoffs. Here’s how to think about them:

Factor	Full Vibe	Careful Vibe	No Vibe
Speed	🚀🚀🚀	🚀🚀	🚀
Risk	⚠️⚠️⚠️	⚠️⚠️	⚠️
Learning	📉	📊	📈
Quality	🎲	🎯	✅
Fun	😎	🙂	😐

There’s no universally correct position on this spectrum. The right choice depends on:

1. What are you building? (Stakes)
2. Who will use it? (Audience)
3. How long will it live? (Lifespan)
4. Who will maintain it? (Future developers)

---

Context Is Currency

The Manning book’s core insight applies here: “Specific beats terrific. High-resolution context = high-precision code.”

When assessing risk, context is everything:

The Same Code, Different Risk

def calculate_total(items):
    return sum(item.price for item in items)

This function is:

• 🟢 Mostly Harmless in a shopping list app for personal use
• 🟡 Caution Advised in an e-commerce prototype
• 🟠 Danger in a production checkout system
• 🔴 Dragons in a financial trading platform

The code is identical. The context determines the risk.

Check the Vibe

Before you vibe code anything, ask:

“If this code is completely wrong, what’s the worst that happens?”

• “I have to restart the script” → 🟢 Vibe away
• “I lose an hour of work” → 🟡 Maybe skim the diff
• “Users see an error” → 🟠 Read carefully
• “Users lose money” → 🔴 Manual mode

---

The Mostly Harmless Checklist

Before going full vibe on any code:

Environment Check

• Is this a personal/learning project?
• Is the data fake or backed up?
• Can I easily revert if something breaks?
• Am I the only person affected by bugs?

Consequence Check

• If this breaks, will anyone lose data?
• If this breaks, will anyone lose money?
• If this breaks, will anyone be harmed?
• Could this create security vulnerabilities?

Scale Check

• How many people will use this?
• How long will this code exist?
• Will someone else need to maintain this?

If you answered “yes” to any consequence question, or your scale is significant, reduce your vibe level accordingly.

---

The Street Rules

Practical wisdom for risk-aware vibe coding:

Street Rule #1: Prototype Wild, Production Careful

When exploring:

"Build me a quick prototype of a user dashboard with
charts showing activity over time. Don't worry about
error handling—just get something visual working."

When shipping:

"Review this dashboard code for production. Check for:
- Proper error handling for API failures
- Input validation on all user inputs
- Performance with 10,000 data points
- Accessibility compliance"

Street Rule #2: Separate Generation from Integration

Generate code freely, but integrate carefully:

1. Generate: Let the AI create in a sandbox
2. Review: Read and understand what it created
3. Test: Verify it works in isolation
4. Integrate: Carefully merge into your codebase
5. Verify: Test the integration

Street Rule #3: Trust But Verify, Proportionally

Risk Level    | Verification Effort
--------------+--------------------
🟢 Mostly Harmless | Run it
🟡 Caution    | Test happy path
🟠 Danger     | Test edge cases
🔴 Dragons    | Prove correctness

Street Rule #4: When in Doubt, Slow Down

The vibes will still be there tomorrow. If you’re unsure about the risk level, treat it as one level higher than you think.

---

Move to Make

After reading this chapter, do the following:

1. Audit your current projects - Classify each as Mostly Harmless, Caution, Danger, or Dragons

2. Adjust your workflow - Match your vibe level to the risk level of what you’re building

3. Create templates - Write prompts for each risk level:
   • “Quick prototype, don’t worry about edge cases”
   • “Production code, include error handling and validation”
   • “Security-critical, I’ll review each line”
4. Practice awareness - Before each AI interaction, consciously identify the risk level

---

The Philosophical Bit

The original Hitchhiker’s Guide rated Earth as “Mostly Harmless” as a kind of cosmic dismissal—a planet of little significance in the grand scheme.

But “Mostly Harmless” is actually a profound compliment for code. It means:

• It probably won’t hurt anyone
• Failures are recoverable
• Experimentation is safe
• You can move fast without fear

The goal of good engineering isn’t to eliminate all risk. It’s to know what risks you’re taking and match your process to those risks.

Full vibe coding on a prototype? Excellent. Full vibe coding on a medical device? Reckless.

The wisdom is knowing the difference.

---

The Guide’s Verdict

Earth: Mostly Harmless.

Your side project: Probably also Mostly Harmless.

Your production database migration: Definitely not Mostly Harmless. Read the diffs.

---

Next: The Total Perspective Vortex