Example draft
A backstory loss, with no cause claimed
Fictional example data; the complete input is included in the package.
Output excerpt
Fictional sample data.
From 2:07.0 to 2:38.5, retention falls 14.2 points, from 64.0% to 49.8%. That is 8.7 times the typical rate of decline in the body of this video. Relative retention falls from 0.55 (above typical for videos of this length) to 0.35 over the same span.
The sheet-building resumes at 2:50.0 ("Okay, back to the sheet"), 48 seconds after the story began. The title promises a budget sheet in 10 minutes; the backstory does not move the sheet forward.
Not established: Why viewers left. The transcript shows what was said, not what was on screen (for example whether the sheet stayed visible during the story) or the audio.
An excerpt from an Agent run using fictional data. Localized text does not imply a separate run. See references/example.md in the package for the full input, output and limits.
How to use
SandBase connection for live research; supplied material can be used offline. Data-call prices and shared free quota depend on the inspected endpoint. Private/owned inputs cannot be reconstructed from public APIs.
Download and unpack the complete Skill folder, keep its references, and use it with a local Agent that supports Skills.
Download Skill v0.2.0
Ask your Agent
Use youtube-retention-diagnosis. Start from the account, link, topic or material I provide, and use SandBase where needed to complete the requested task. Deliver: A calculated retention diagnosis and one bounded change to test