Can you tell which answer had the skill?
Two AI answers to the same task: one written with the handoff skill, one with unrelated text of the same length in its place. Guess which is which.
Try every skill, both models, and the fake vs nothing twist
Report 017 gave an AI model the same tasks with a skill, and again with a placebo in the skill's place. The placebo is unrelated text cut to the same length as the skill's instruction text. Same length means the instructions given to the model, not its answers.
Each pair below is one task from that run. One answer came from the model given the handoff skill. The other came from the same model given the placebo instead, on the same try of that task, matched by draw number. They are shown in random order as Answer A and Answer B. Pick the one you think had the skill, then press Reveal.
This is a game of noticing. It is not a test of you, and not a measurement of the skill.
One reader's picks on these pairs. It is not a measurement of the file: it moves no verdict, and it says nothing about other readers, tasks or models.
Every answer written with the skill has a "Suggested skills" section, which the skill's text asks for. Most open by saying they did not save the file the skill asks for, and most run longer than the placebo answer beside them. A right pick can come from those cues, not from the skill helping.
This page uses handoff: the shortest skill counted in Report 017's headline whose text does more than hand off to another skill, chosen by its text and length alone, not by any result. grill-me and grill-with-docs are shorter, but each one's whole text only tells the model to call other skills.
Report 017 reads handoff against its placebo as "case rule: at least one case scored below the same-length placebo". In plain words, on at least one task the skill's answers scored clearly below the placebo's: beyond the spread between tries and by at least the report's margin.
These are 12 tries of 3 text-only tasks, answered by Claude Sonnet 5.5 and scored by Claude Haiku 5.5 on 1 November 2026. The tasks forbid saving files, which handoff asks for. This reads this run, not handoff in general.
12 of 12 verified pairs. Dropped 0 pairs.
Picking at random, a reader gets 10 or more of these 12 right about 1.9% of the time.
The page source holds which answer is which. Read the pairs before you read the source.
The pairs
What stays in this page
Your picks never leave this page: it sends none of them, stores nothing and calls no model. The site's visit counter (Cloudflare Web Analytics) still counts the page view.
Where the pairs come from
Receipt c99347fd5003d4ce8f8af82defa7c3480943b1d348dc35c73845f59cf9776258. Each answer is shown whole, and its SHA-256 is the one the receipt seals for its task, arm and draw. Each prompt is the task of the suite whose hash the receipt seals.