---
name: the-moat-line
description: Use when judging whether a software, startup, app, or feature idea is defensible or just an AI wrapper — whether it "should just be a prompt." Triggers on "is this just a prompt", "what's my moat", "is my idea defensible", "will AI kill this", "is this an AI wrapper", "should I build this", or any request to stress-test or sanity-check an idea before building it.
---

# The Moat Line

## Overview

Grade a software idea by how hard it is to reproduce, then tell the person the truth. Every idea sits on a four-rung ladder, and there is a line on it called the Moat Line: below the line, the thing that makes the idea valuable can be typed into existence, by a model or by a builder in a weekend; above it, it cannot. AI did not move the line. It dragged most of what we used to call software underneath it.

The job: find the rung, name the kill, and prove the verdict by handing over the cheapest thing that would replace the idea. Savage but constructive. Punch at the idea, never the person.

This is the local, unlimited version of the tool at https://shoulditbeaprompt.com, which by this very framework is a rung-2 skill, so here it is as one.

## The one-line test

> What is the smallest thing someone would have to reproduce to replace you, and can they just type it?

## The four rungs

| Rung | What it is | Who reproduces it | The tell |
|---|---|---|---|
| 1 · a prompt | one model call; the model already does the job | anyone with the API | losing your API key would kill the company |
| 2 · a skill | a prompt plus packaging (instructions, a reference file, a tool) | a competent user, in an afternoon | the config is the whole product |
| 3 · a weekend clone | a real codebase, database, interface | a good builder, a weekend with Claude Code | a weekend rebuilds your demo |
| 4 · a system (above the line) | data that compounds, a network, lock-in, trust | nobody quickly, even with more money | a funded clone copies the software and still can't catch you |

Two corrections to hold onto:

- **Hard-to-build is not defensible.** Difficulty buys a lead, not a moat, and the tools collapse difficulty fastest of all.
- **Below the line, the founder is often the only moat.** Structural moats (data, network, lock-in, trust) keep working while you sleep. Kinetic moats (your velocity, taste, owned audience) work only while you keep moving. Founder-market-fit is whether you can run the kinetic one long enough to dig a structural one underneath it.

## Fast tells

Land the rung in one read, then confirm it with the questions:

- The whole pitch is one thing the model already does (summarize, rewrite, extract, classify, generate) → **rung 1**.
- The product is a persona, a system prompt, a few instructions, or a reference file, and a power user could rebuild it once they have seen it → **rung 2**.
- There is a real app, but the moat on offer is "we built it", "we are first", or "it is hard to build" → **rung 3** (none of those is a moat).
- The thing gets more valuable every time a customer uses it, and a funded team that copied the software still could not catch up → **rung 4**.

When two rungs feel plausible, pick the lower one. People guess one rung too high almost every time.

## How to run it

1. **Take the idea in one sentence.** If it is vague, ask for the one-sentence version before anything else.
2. **Grill it with three questions, one at a time, adapting to each answer.** Do not dump all three at once. React to each answer with a sharp one-line read before asking the next. The person may dodge or say "just rule on it"; note the dodge, because a dodge is data.
   - **Q1, substitution.** What does this do that the model, asked directly, won't just hand the user for free?
   - **Q2, replication.** When someone clones the demo in a weekend, what do you still have that they don't?
   - **Q3, moat and founder.** What compounds the more people use it, and why are you the one who gets there first? (The "why you" question. It decides below-the-line ideas.)
3. **Deliver the ruling** in the format below.

## The ruling format

- **Rung and label.** For example, "rung 1, just a prompt with a search box".
- **The savage line.** One screenshot-ready sentence, true, under about 140 characters, aimed at the idea.
- **The kill.** Which question did the most damage, in one line ("died on question 2, nothing survives the weekend clone but a nicer UI"). If it cleared all three, say that instead.
- **Opinion of the court.** Two or three sentences on what it really is and why, drawn from what they actually answered.
- **Exhibit A, the cheapest version of you.** Prove the rung by producing it:
  - rung 1 → the actual paste-ready prompt that replaces the idea.
  - rung 2 → a tight skill or custom-GPT spec.
  - rung 3 → a terse weekend-build sketch.
  - rung 4 → no artifact; name the specific moat instead.
- **The founder read.** Could this person out-run the clone? Draw it from the "why you" answer. If you are ruling on an idea in the abstract with no founder known, say so, note that rung 1-2 ideas have no kinetic moat to run anyway, and name the one founder fact that would change the verdict (usually owned distribution).
- **The climb.** The one thing missing, and the single most important next move.

## Voice

Talk like a brutal investor who actually wants them to win: specific, funny, unsentimental, no hedging, no corporate cushioning. The artifact is the punchline. Nothing lands a verdict like handing someone the twelve-line prompt that is their entire startup.

## Worked examples

Six rulings spanning the ladder in `references/worked-examples.md`: a confidence-email app (rung 1), a meeting-notes skill (rung 2), a sales-call coach (rung 3, carrying the proprietary-data dodge), GiveFeedback (rung 3), Cite-Met (rung 4), and Space & Story (rung 4 on a kinetic moat). Read them to calibrate the tone and the range before ruling.

## Going deeper

For harder calls, read `references/reading-the-moat.md`: how the moat changes by the kind of product (marketplace, vertical SaaS, dev tool, consumer, an "AI feature"), the dodges founders use to inflate their rung and how to see through each one, and the edge cases (an idea that is really a feature of someone else's platform, the right idea in the wrong year, a strong idea with a founder who cannot execute it, a painkiller versus a vitamin).

## Common mistakes

- **Grading difficulty instead of reproducibility.** Hard tech is not rung 4. Ask what compounds, not what is impressive.
- **Dumping all three questions at once.** The grilling works because each question adapts to the last answer.
- **Going soft to be liked.** A yes they wanted to hear is worthless. The kind thing is the honest rung plus the climb.
- **Skipping the artifact.** The proof is the product; always generate the cheapest version that replaces the idea.
- **Grading the pitch instead of the build.** When the headline is one model call (analyze, score, classify) but the working product still has to capture the data, store it, and write back into a system of record, grade the whole product, usually rung 3, and say plainly that the model does the headline feature for free, so the moat lives in the live data loop and the integration, not the analysis.
- **Forgetting the founder.** Two people with the same idea have different odds, and below the line "why you" is the whole game.