You've probably heard someone say "AI is like having a really smart intern." It's a popular metaphor. It's also why you keep getting frustrated.
When you treat AI like an intern, you expect it to understand what you mean. You rewrite your instructions hoping this time it'll get it. You feel like you need to motivate it, explain context, maybe even be encouraging.
That's the wrong mental model.
Andrej Karpathy recently described AI as more like a "ghost" — something powerful but fundamentally alien. It doesn't understand your intent. It doesn't learn from your tone. It pattern-matches. And once you internalize that, everything about working with AI clicks into place.
Here's what the "ghosts, not employees" model means for your business.
The Problem with the Employee Metaphor
When you think of AI as an employee, you bring a ton of assumptions:
- "If I explain it better, it'll understand." So you write longer prompts, add more context, rephrase things three times. The output gets worse, not better — because you're adding noise, not clarity.
- "It should remember what I told it before." So you get annoyed when it forgets your business context between conversations. But it doesn't "forget" — it never "knew" in the first place.
- "It should care about quality." So you're surprised when it produces sloppy output. But it doesn't care about anything. It predicts likely next tokens based on patterns.
The employee metaphor creates emotional friction. You feel like you're managing someone who won't listen. But you're not managing a person — you're configuring a tool.
The Ghost Model: What Actually Works
Think of AI as a ghost: powerful, fast, tireless, but operating on alien logic. It doesn't share your assumptions. It doesn't infer your intent. It does exactly what you configure it to do — no more, no less.
This model changes your behavior in three practical ways:
1. Replace "Explain Better" with "Verify Automatically"
Stop trying to write the perfect prompt. Instead, build verification into your process.
Bad approach: Write a 500-word prompt explaining your brand voice, then hope the AI nails it.
Better approach: Give it a short, clear instruction. Then check the output against a simple rubric. If it fails, adjust the instruction — don't add more paragraphs.
Best approach for repeatable tasks: Set up an AI loop. The AI generates, then a second pass evaluates against your criteria, and they iterate until the output meets your standard. This is how you get consistent quality without babysitting.
Example: Instead of "Write a professional follow-up email that sounds friendly but not pushy" (vague, subjective), try "Write a follow-up email. It must: include the client's name, reference their last service date, offer exactly one next step, and be under 100 words." (specific, verifiable).
2. Give It Metrics, Not Vibes
Ghosts don't read the room. They read instructions.
If you want good output, replace subjective language with measurable criteria:
| Vague (vibes) | Specific (metrics) |
|---|---|
| "Make it professional" | "Use no slang, no exclamation marks, formal salutation" |
| "Keep it short" | "Maximum 150 words" |
| "Sound like our brand" | "Use these 5 phrases; avoid these 3 phrases" |
| "Make it persuasive" | "Include one statistic, one testimonial quote, one CTA" |
The more your instructions look like a checklist, the better AI performs. Not because it "understands" checklists — but because checklists are unambiguous pattern-matching targets.
3. Configure, Don't Convince
Here's the shift that saves the most time: stop convincing AI and start configuring it.
Convincing looks like: "You are an expert copywriter with 20 years of experience in the HVAC industry. You understand that homeowners are worried about energy costs and want reassurance. Please write a blog post that..."
Configuring looks like: "Write a 600-word blog post about HVAC energy efficiency. Target audience: homeowners. Include: 3 actionable tips, 1 call to action for a free estimate. Tone: informative, not salesy."
The first prompt wastes tokens on role-play that doesn't improve output. The second gives the AI clear constraints it can actually follow. Same result, less frustration, lower API cost.
What This Looks Like in Practice
Let's say you run a landscaping business and you want AI to write your weekly newsletter.
Employee-mindset approach: Spend 20 minutes crafting the perfect prompt about your business, your tone, your audience, your goals. Get a draft that's "okay." Tweak the prompt. Try again. Still not quite right. Add more context. Now the prompt is longer than the newsletter.
Ghost-mindset approach: Create a simple template. "Write a 300-word weekly landscaping newsletter. Include: one seasonal lawn care tip, one promotional offer, one FAQ answer. Use short paragraphs. No jargon." Run it. Check the output against your 3-point checklist. If it passes, send it. If not, adjust one constraint and re-run.
The ghost approach takes 5 minutes. The employee approach takes 45 minutes and produces worse results.
The Bigger Picture
The businesses getting real value from AI right now aren't the ones writing the most sophisticated prompts. They're the ones who've stopped treating AI like a person and started treating it like a configurable tool.
They've built simple processes: generate → verify → ship. They've replaced subjective feedback with objective criteria. They've stopped hoping AI will "just get it" and started giving it inputs that produce predictable outputs.
Your AI doesn't need motivation. It needs configuration.
Stop prompting. Start looping. The ghosts work faster when you give them clear paths instead of open roads.
If you're running a small business and want to see what configured AI workflows actually look like — not hypothetical, but the specific prompts and loops that save hours each week — follow along at SMB Scale Up.
Top comments (0)