The Door Rule: A Mental Model for Prompting

Every bad prompt has the same root cause

You've sent a prompt to an AI model and gotten back something useless. Generic advice. A response that misses the point entirely. Maybe the model asked you three clarifying questions before doing anything at all.

The instinct is to blame the model. "It didn't understand what I wanted." But the model performed exactly as instructed. The problem was the instruction itself.

There's a mental model that fixes this. It's attributed to Andrej Karpathy, one of the most respected figures in AI, and it reframes how you think about every single prompt you write. We call it the Door Rule.

What the Door Rule actually is

Picture this. You need someone to complete a task, but that person is in another room. The door is closed. The only way you can communicate is by writing on a piece of paper and sliding it under the door.

You never see each other. You can't gesture. You can't point at your screen. You can't say "you know, that thing from last week." You write your instructions on paper, slide it through, and wait.

The person on the other side reads your paper and does whatever it says. Then they write their response on paper and slide it back.

That is exactly how prompting works. What you write on that paper is your prompt. What comes back is the model's response. The model cannot see your screen. It doesn't remember your last conversation. It doesn't know what you're thinking about. It doesn't have access to files you haven't provided. It has nothing except the words on that piece of paper.

Here is what this looks like when it goes wrong:

Bad paper (sliding under the door):

What do you think about this?

> The person on the other side has no idea what "this" refers to. They can't see your screen. They don't know if you're talking about a spreadsheet, an email, a business plan, or a sandwich. All they can do is ask you to clarify, which is exactly what the model will do.

Another bad paper:

Please analyze this email.

> There's no email on the paper. The person in the other room has nothing to analyze. The model will either make something up or ask you what email you're talking about.

Good paper (sliding under the door):

Here is an email from a potential client asking about our Q2 availability:

"Hi, we're looking to bring on a fractional COO for Q2. Are you available?
Our budget is $15K/month. We're a 40-person SaaS company."

Analyze this email for: buying signals, red flags,
and a recommended response strategy.

> The person on the other side has everything they need. The email itself, the context around it, and specific instructions on what to produce. They can get to work immediately.

The difference between those prompts isn't talent. It's awareness. The Door Rule trains you to check your work before you slide that paper through.

The Door Rule checklist

Before you send any prompt, run through this mental checklist. It takes about ten seconds and will save you entire rounds of back-and-forth.

If you're unsure whether you've included enough, add more. Overloading context is almost always better than underloading it. A detailed prompt that's slightly long will outperform a short prompt that leaves the model guessing.

This checklist becomes second nature with practice. You won't run through it every time you fire off a quick question. But for anything tied to a client deliverable, a report, or code generation, it's the difference between a usable first draft and a wasted turn.

> The real skill shift. When a model gives you a poor response, stop blaming the model. Go back and inspect the paper you slid under the door. Almost every time, you'll find missing context, vague instructions, or assumptions the model had no way of knowing. The fix is always in the prompt.

What this looks like in operator workflows

The Door Rule matters most when the stakes are real. For fractional leaders managing multiple client engagements, every prompt is a piece of paper sliding under a door to someone who has never met your client, never seen your data, and has no memory of your last session.

Client deliverables. When you're drafting a board update for a client, the model doesn't know the company's metrics, the audience's expectations, or your preferred structure. A vague prompt like "write a board update" produces something generic. A Door Rule prompt includes the actual numbers, the key narratives, the format from last quarter, and the tone the board expects.

Code generation. This is where the Door Rule hits hardest. Asking a model to "fix this bug" without providing the relevant files is like asking a developer to patch a codebase they've never seen. They can't do it. You need to include the specific files, the error message, what you expected to happen, and what actually happened.

Multi-client context switching. When you're jumping between three engagements in a single afternoon, it's tempting to fire off quick prompts. But the model resets between conversations. It has no idea you were working on Client A's pipeline ten minutes ago. Every new prompt is a fresh piece of paper to a person who has never heard of you.

Here is a before-and-after for a typical operator task:

Without the Door Rule:

Create a weekly status update for my client.

With the Door Rule applied:

Create a weekly status update for Acme Corp (Series A SaaS, 40 employees).

Key updates this week:
- Completed migration from Airtable to HubSpot CRM
- Two new enterprise leads entered pipeline ($50K+ ARR each)
- Engineering shipped the billing dashboard, one day behind schedule

Format: 3-5 bullet points with a one-sentence executive summary at the top.
Tone: Professional, direct, no filler. This goes to the CEO and VP of Sales.

The second prompt takes 60 extra seconds to write. The response comes back ready to send, instead of requiring two more rounds of corrections.

Make this automatic before your next prompt

The Door Rule isn't a technique you use sometimes. It's a filter that runs in the background every time you interact with a model. The more you practice, the less you think about it. Your prompts get more specific by default. Your first-draft responses get more usable. You spend less time going back and forth.

Here is the action step. Pick one prompt you plan to send today -- a real one, tied to actual work. Before you send it, pause and ask: "If I slid this paper under a door to someone who knows nothing about my situation, would they have everything they need?"

If the answer is no, add what's missing. Include the data. Spell out the format. Define what "good" looks like.

That single habit will do more for your prompt quality than any collection of templates or pre-written frameworks.

> The pattern to internalize. Bad results are almost never a model problem. They're an instruction problem. The Door Rule gives you a repeatable way to catch that before you hit send, not after you've already wasted a turn and started over.