Skill Lift

Skill Lift is the score difference you get when you make the same AI agent solve the same problem twice, once with a manual and once without. Think of someone new to a neighborhood wandering without a map versus someone walking around with one. Same person, same destination, but having that one sheet of paper changes how long it takes and how many wrong turns they make.

The "manual" here is what's called an agent skill: a document prepared in advance that explains what a program does, when to call it, and how to call it, which is then handed to the agent. If you keep the prompt, the model, and the task completely fixed and only change whether this manual is present, the resulting score difference is that skill's real contribution — in other words, its Skill Lift.

This number matters because it separates skills that merely sound good from ones that actually help. No matter how convincingly a manual is written, if attaching it doesn't raise the score, it's useless. But if a measurable improvement shows up, that gives you a real basis for trusting and using it.

Articles cite Skill Lift figures like this: "accuracy in getting the right answer rose by 41 points, effectiveness in actually achieving the user's goal rose by 39 points, and efficiency in reaching the goal without unnecessary steps rose by 35 points." A common misunderstanding is treating this number as an improvement in the model itself, but it's really just the difference produced by attaching one manual while leaving the model untouched. In other words, it doesn't make the model smarter — it helps the model make better use of abilities it already had.

To get a feel for Skill Lift yourself, try giving an AI tool an unfamiliar task with no explanation, and count how many attempts it takes to get the result you want. Then summarize how to use that same tool in one paragraph, put it at the top of your prompt, and make the same request again. The difference in the number of trial-and-error attempts between the two runs is that manual's Skill Lift.