How Rate My Clip Actually Works

In the last post, we argued that the single-prompt virality scores most clip tools ship are compressed into a useless 72-to-94 band. The fix isn't “ask a better prompt.” It's to change the architecture. This post walks through the decisions behind Rate My Clip, at a level high enough to explain the why without giving away every knob.

The Shape of the Pipeline

When you paste a YouTube Shorts URL and hit submit, the clip goes through four rough stages:

1. Prep. We grab the video title and thumbnail. Nothing fancy.
2. Grading. A small ensemble of judges watch the clip in parallel. Each one grades the same rubric, independently.
3. Reconciliation. The individual verdicts are combined into a single per-dimension tier by majority vote, along with a confidence signal based on how much the judges agreed.
4. Editing. A final pass merges the judges' raw notes into one polished version per dimension, with a short punchy summary for the share card.

End-to-end it takes about 30 seconds, most of which is the judges watching the video. The rest of this post is about why the pipeline looks like that and not like one call to one model.

Why More Than One Judge

The single-prompt problem we described last post has two roots: models hedge toward safe middle numbers, and whatever felt vaguely right in one pass becomes the final output. Running the same rubric a few times doesn't magically fix hedging on its own, but it gives you something the single-pass version can't: disagreement as signal.

On each dimension, one of a few things happens:

• Unanimous. All judges landed on the same tier. High confidence.
• Majority. Most agreed, one dissented. Medium confidence.
• Split. Nobody agreed. Low confidence. The dimension is genuinely ambiguous for this clip.

That confidence level is shown in the report. “Most of the judges said your payoff was weak, one said it was average” is actually useful feedback. “Viral Score: 78” is not.

Calibration With Anchors

The single biggest thing that keeps tiers from collapsing into a generic middle is anchoring. Before any judge grades your clip, it's shown a small set of reference shorts whose quality is already known: a clearly good one, a middling one, and a clearly bad one. Each reference comes with per-dimension notes so the judge knows why the reference sits where it does.

Then, for each of the rubric's questions, the judge is asked a concrete comparative question: is the target closer to the good reference, the middle one, or the bad one? That framing matters. “Rate this hook 1 to 100” is abstract and invites hedging. “Is this hook closer to this known-good example or this known-bad example” is concrete and forces commitment. Calibration through example, not through numbers.

Tiers Instead of 1-100

The judges never output a number between 1 and 100. They output a tier per dimension: BAD, AVERAGE, or GOOD. Each tier has a specific textual definition of what qualifies. The 0-100 score you see on the report is computed after, by weighting each dimension's tier.

The reason for this is the entire thesis of the last post. You can't hedge into the 80s if “the 80s” isn't an option. You can only pick BAD, AVERAGE, or GOOD, each of which corresponds to something you have to justify.

This is also why BAD actually gets used. The judge prompt includes an explicit anti-hedge rule: don't default to AVERAGE as a safe middle ground. If the target is clearly bad on a dimension, say so. Combined with the tier-structured output, judges actually give weak hooks weak tiers. That's how the system can honestly tell you your hook isn't working.

Reasoning Before Grading

One quieter detail: judges are required to describe and justify before they commit to a tier. For each dimension, they produce an observation about the target video first, then a short critique comparing it to the references, and only then a tier. Forcing the order this way reduces the “pick an answer, rationalize after” failure mode that pure 1-to-100 prompts fall into. The tier becomes the consequence of the reasoning, not the other way around.

The Final Editor Pass

After the ensemble finishes, we have several complete drafts per dimension. Good drafts, but overlapping and occasionally contradicting. Showing all of them would be noise. Picking one would waste the others.

So we run one more pass. This one isn't a judge. It's an editor, and the tiers are already locked by majority vote before it sees anything. Its job is to merge the drafts into one polished observation, one critique, one concrete action, and a short snappy summary per dimension. The summary shows up on the share card and near the top of the report.

The editor can't change a tier. If it tries to hedge, the tier still wins. If the editor pass itself fails for any reason, the report falls back to showing a judge's raw draft. Never a blank report.

Why It's Free

Rate My Clip is cheap enough per grade that we can run it without a login and without metering. If a single grade cost a lot more, we'd have to gate it behind signup or credits. It doesn't, so we don't.

The deeper reason, though, is that it's a sharper demo than any landing page. If you use the grader and recognize the feedback as honest, “yeah, my hook is weak,” that's a better argument for the main product than anything we could write in marketing copy. ClipFinder runs the same philosophy on the harder “find the good clips” problem: multiple specialized passes, structured outputs, anti-hedge rules, no fake scoring theater.

Try It

Paste any YouTube Shorts URL here and you'll get the full report in about 30 seconds. Multiple judges, six dimensions, tier votes, confidence, summaries, concrete actions. Free, no signup, no credit card.

If the feedback reads as real, and not like a mood ring, that's the architecture doing its job.