Vibe Coding
Build a small multi-page web app from a written spec, including one form and one piece of saved data. Record whether it ran, how long it took, what it cost, and how much a new person had to already know.
Method
How we rank
A number appears only after a tool is run on a published task. Until then the listing says pending, and the leaderboard is alphabetical.
The four scores
- Quality
- Did the bench task produce something a builder could keep? We look at the output, not the demo video.
- Price
- What it actually cost to finish the task, including usage beyond the monthly plan. A cheap plan that cannot finish the task scores poorly.
- Speed
- Clock time from a clean start to a usable result. Waiting on a queue counts.
- Ease of starting
- How much a new person had to already know. Setup rabbit holes are part of the score.
Overall is the arithmetic mean of the four, rounded to the nearest integer. Ties break by name, A to Z. The content schema rejects a tested record whose overall does not match that mean, so a hand-edited score cannot drift from the rubric.
The benches
AI Agents
Complete a multi-step job that calls one external tool and produces a checkable result. Record reliability, cost, time, and how long setup took from a clean machine.
Game Dev
Run the job the tool claims: a small sprite sheet in a consistent style, a game-ready 3D prop, a spoken NPC line, or a one-minute playable. Judge the output a developer could actually drop into a prototype.
Ship & Grow
From a clean account, finish one task a builder would do in a day: deploy an app, store a row, take a test payment, send a transactional email, or publish a doc page. Record time, cost, and the steps that blocked a new person.
Cadence and money
A tested tool is due for a re-test at least monthly, and sooner when the vendor ships a major version. The tool page shows the last tested date. A stale date is a fact, not a hidden penalty, until the re-test replaces the score.
We do not sell rank. An affiliate link is labeled and points at a separate button. A sponsor can buy a newsletter mention, a Featured badge, or a faster place in the test queue. None of those fields are read by the sort. The code is in the repository. The disclosure page defines each label.
Changes to this method are reviewed by a person before they go live. A new weight, a new bench task, or a dropped criterion is a change to the method. Adding a tool is not.
As of launch, every score is pending. If you see a number, it should match a dated test note. If it does not, that is a bug.