We test every tool the same way, so you can trust the comparison.

Before any tool is tested, the editors write a category brief: what the tools in this category are actually trying to do, what a good outcome looks like, and what the category edge cases are. The AI coding assistant brief looks nothing like the voice AI brief — and that's the point.

The brief defines the prompt battery: real-world use cases the graders have actually hit in their own work — engineering, writing, research. Every task is drafted from something someone needed to do for real, then tightened to stress a specific dimension. We retire prompts as tools start to ace them and add new ones as the category moves.

Every tool in a category faces the same prompt set. Same wording. Same seed where supported. Two reviewers run them independently to catch flukes. Outputs go straight into the graded log.

Sample prompts shown above. The full battery is kept internal for now — every prompt, temperature, seed, and pass criterion is logged against the same rubric for every tool. We'll open up the battery and its logs category by category as each one stabilises, so readers can inspect and reproduce what we ran.

Batteries catch raw capability. Daily use catches everything else: friction, fatigue, the small UX papercuts that make a tool quietly get ignored by week two.

Each tool gets plugged into a real workflow — typically a live team's Slack, Notion, or codebase — and left there for two weeks. The reviewer journals every session. At the end, the team answers one question: did you keep using it voluntarily?

Every prompt output is graded independently by two reviewers per tool on a 1–10 rubric. An AI panel (Claude alongside peer models, same rubric) runs in parallel as a calibration check — not as a vote.

If the two reviewers disagree by 2+ points, a third breaks the tie and we flag the prompt for battery review. We publish the full disagreement rate with every scorecard. Calibration is an editorial KPI.

In parallel with the lab work, our signals pipeline pulls fresh discussion about each tool across the open web. Claude structures the raw text into sentiment, top strengths, most-complained-about issues, and pricing moves.

Live signals don't move the RAI score — they're shown next to it. If the web wildly disagrees with our verdict for more than a week, the editorial team triggers a re-review.

Six dimensions. The weights below are our defaults — what we think matters most for a general reader. On the browse page, you can slide the weights to your own use case and the catalog re-ranks live.

The RAI score is a weighted sum. No magic, no secret sauce, no vendor-specific bumps. Sliders on the browse page recompute it live with your own weights.

A review is never “done.” We re-run the battery on every major vendor release, and at minimum every 90 days. Scores move. Verdicts get rewritten. The per-tool changelog records every edit, with reasons.