Methodology

How we check prompts. How we score tools.

01 · Prompts

How we check prompts

Prompts go stale. A vendor updates a model and the prompt that worked great last month starts slipping. The answers get weaker and you assume it's you. It's not. We curate the high-impact prompts and re-check every one, every month. The one you copy is as good as we can make it, today.

Nobody warns you when a prompt stops working. It still runs. Like a watch that's slow and getting slower. Still ticking, still wrong. And you're two hours late to the meeting with your biggest customer. They don't blame the watch.

  • One prompt per model. Not one prompt for everybody. We write each job separately for each model, because the same words land differently on different machines. You grab the one built for the AI you already use. No rewording.
  • Read first. Dated second. Every month we read every version in full against that vendor's own current prompting guide, and fix what no longer holds. That's a read, not a test drive. The vendors keep changing the rules. Keeping up is our job, not yours. The date is the day of that check. Not the day it was written. Not the day the page was built.
  • No date it didn't earn. If a prompt hasn't been checked yet, the page says NOT YET CHECKED, right there on the prompt. We'd rather show a gap than a number. So when you do see a date, you can lean on it.
  • Models change. The prompts keep up. New major model, new prompts. Small update, same prompts. When a vendor pulls a model, its prompts come down, so you don't end up pasting something built for a model you can't open anymore.
  • Every check leaves a receipt. Which model got checked. How many live variants it covered. Any date on any page traces back to the check that put it there. You don't have to take our word for it.

The date at the top of the prompts page is the OLDEST of those check dates, never the newest. If one model was re-checked yesterday and another three weeks ago, the bar shows three weeks ago. A claim about the whole library has to be true of its weakest part.

02 · Tools

How we score tools

We score what works. Every score on Unbought comes from real, first-hand user sentiment: what actual users report after putting a tool to work. Not press releases, not paid placements, not our opinions. One number, and it can't be bought.

One number, the Unbought Score

Every scored tool gets one number from 0 to 100, its Unbought Score. It's built only from verified first-hand user experience and adjusted for how much real evidence stands behind it, so a tool with a handful of glowing mentions can't leapfrog one proven across hundreds of real reads. When there isn't enough signal to score a tool honestly, we don't guess: it shows AWAITING DATA instead of a number.

The 6-layer pipeline

We call it Community Intelligence, and it runs in six separate, auditable layers. Every layer is measured against human-labeled validation sets before it ships.

  1. Verbatim capture. Every review is captured word-for-word and never edited. The raw record is immutable.
  2. Normalization. Captured text is cleaned and standardized the same way every time, so the same input always produces the same result.
  3. Attribution. A 12-signal engine binds each comment to the specific tool it's actually about: no guessing, no lumping similar tools together.
  4. First-hand verification. Each comment is independently checked against one standard: did this person actually use the tool themselves? Recommendations, hearsay, and marketing noise are thrown out before anything is scored.
  5. Human-anchored QA. Results are checked against labeled gold-standard sets graded by people, so the machine's judgment stays honest.
  6. Sentiment and score. Only the verified first-hand experiences are read for sentiment and aggregated into the final Unbought Score.

More first-hand reads means higher confidence, which we show as a confidence band alongside the score.

Verdicts

WORKED (75 and up) · MIXED (50–74) · FAILED (below 50). The verdict is the score, translated, never edited by hand.

The rules we don't break

Zero affiliate links. Zero sponsored placements. Zero pay-to-influence rankings. Collections are curated by humans, but the ORDER inside every list always comes from the scores, never from us, never from a sponsor. User quotes appear verbatim or not at all. Corpus-level totals (like the counts on our homepage) are shown live from the database; we don't decorate individual tools with raw counts.