Free preview8 min

The Experiment Setup

Reviewed by Human · Updated October 19, 2026

You give a capable model one clear sentence: build an app that shows documents and answers questions. It writes confident code for twenty minutes — and hands you something that won't even open.

Live Demo — AI Terminal

Watch a prompt-only run of Project 01's task. The starter folder has no rules and no checks — just the app skeleton and one vague prompt.

Your prompt

Your turn

Before you read on, predict what a prompt-only agent will get wrong. Project 01 wants four features: window startup, document list, QA panel, and a data directory. Which one will a single vague prompt most likely fake or skip?

Reflect

Most people guess the QA panel gets faked — it's the easiest to stub with code that looks right and does nothing.

What Happens When You Run an Agent With Just a Prompt?

**Baseline run**: an agent given only a task prompt, with no rules, state, or checks around it. In Project 01 you run this first, on purpose, to see the failure clearly before you fix it. Three things tend to break. The agent claims 'done' with no proof. It half-finishes features because nothing tracks which are actually complete. And it can't tell working code from plausible-looking code, because there's no check to run. This is the same lesson from why strong models still fail and what a harness actually is: the gap isn't intelligence, it's the environment. You're about to watch that gap on one real task.

A baseline run isn't a mistake — it's your control group. You can't measure a harness without one.

In a prompt-only baseline run, why does the agent so often say 'done' when the app doesn't work?

📚

Go deeper

end of module

You've finished this module.

Mark it complete to earn your XP and keep your streak alive.

Progress saved locally · Sign up to earn XP

Building the Minimal Harness→