Free preview13 min

Compare and Reflect

Reviewed by Human · Updated October 19, 2026

Same task, same model, measured twice

✗ Baseline run (prompt only)

Build an Electron app that can show documents and answer questions.

One vague sentence. No definition of done, no state, no check. The agent finished fast and claimed success, but two of four features were faked and the app didn't open.

✓ Harnessed run (AGENTS.md + init.sh + feature_list.json)

# AGENTS.md ## Definition of done All four features pass verification: - window startup - document list - QA panel - data directory ## Before you start Run ./init.sh and read feature_list.json. Do not mark a feature done until its check passes.

Same task, wrapped in three files. The agent ran init.sh first, worked one feature at a time, and marked done only after a check passed. Slower — and it actually finished.

Your turn

Run — or read — both versions of Project 01 and log the difference on the three axes the project itself uses: completion, retries, and premature 'done.' Numbers beat impressions.

Reflect

If your two columns look nearly identical, your harness probably restated the README instead of adding a check — the exact trap from the last module.

How Do You Measure If a Harness Actually Helped?

**Ablation**: removing or adding one part of a system to measure that part's effect. Project 01 is a two-point ablation — harness off, harness on — on one task. Measure three things, not vibes. Completion: how many of the four features actually pass. Retries: how many restarts or corrections the run needed. Premature done: did the agent claim success before proof. The harnessed run is usually slower and costs more — and still wins, because 'fast and broken' isn't finished. You saw the same theme in why strong models still fail: capability was never the bottleneck. Keep this habit as you move toward making the repo the system of record.

Live Demo — AI Terminal

Watch the moment the harnessed run catches itself — the behavior a prompt-only run never shows.

Your prompt

🔁

The harness is a control group you can keep

Every task you run is secretly an experiment. A baseline run tells you what the model does alone; a harnessed run tells you what your files are worth. Keep measuring both, at least early on, and you'll never again argue about whether a harness helps — you'll have the two columns.

The run that talked itself out of a mistake

▎

📚

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