Send us your skills. We'll tell you which ones you can delete on Claude 5.

Claude 5 models do a lot without help. Some skills still earn their place. Others now do no better than a block of filler text of the same length. Send us a public skill pack, or a single skill, and we'll measure every skill and publish what we find. It's free.

What we do

For each skill we write three tasks from the skill's own text and its README: a normal case, an edge case and a constraint case. We run each task on Claude Haiku 5.5 and Claude Sonnet 5.5 three ways:

• with the skill;
• with no skill;
• with a placebo: a text the same length as the skill that tells the model nothing useful.

A separate grader, Claude Haiku 5.5, scores every answer. The tasks are published with the result, so anyone can check them.

What you get, per skill and per model

• Keep. On at least one task the skill scored above the placebo, and on none did it score below. Its words do something that the same amount of text does not.
• Delete candidate. The skill did not beat the placebo on these tasks: it scored below it on at least one task (even if it scored above it on another), or no task differed by as much as the test can detect. That is a reason to try working without it. It is not proof the skill does nothing, because your own tasks may differ from ours.
• Cannot tell. The results were too uncertain to read either way, or the skill cannot be measured this way.

A skill can be a keep on one model and a delete candidate on the other. We show both.

We measure text-only skills. A skill that routes to other skills or drives tools is marked "not measurable this way", because a text task can't show what it does.

Three ways to ask

1. A pack (up to 20 skills): open a Measure a pack request with the repository link.
2. One skill: open a Measure a skill request.
3. Challenge a result: if you think we tested your skill unfairly, open a Challenge a result request with your own three tasks. We re-run with your tasks and publish both results side by side.

If the skill is private, email hello@driftproofhq.com instead of opening a public request.

The terms

• Free.
• Anyone can ask, not only the author. If you're not the author, the result page says so.
• We publish the result, whatever it says.
• Results come within three days of your request reaching the front of the queue. The queue is below.
• We re-run every result on the next Claude model.
• Open to the first 50 requests. A pack can hold up to 20 skills.

The queue