Our review
This is a fairly substantial workflow for building or improving a SKILL.md skill. It asks the author to define intent, write instructions, create test cases, compare behavior with and without the skill, and use feedback to revise it. References, grading prompts, an evaluation viewer, and packaging utilities support a repeatable iteration loop rather than a one-pass draft.
Use it when a skill will be reused and you can afford to examine actual agent behavior. Skip it for a one-off instruction that does not need formal evaluation. Its own process has many moving parts; a small team may prefer to apply the core idea manually before adopting the full benchmark and viewer setup. Treat evaluation numbers as evidence about the chosen test cases, not a guarantee of future performance.
The folder contains Python scripts, an HTML viewer, and agent prompts. Some steps dispatch agent runs and can consume model usage; local scripts package files and aggregate results. Read the scripts and the current SKILL.md before execution, especially where tests invoke an agent or open a browser.
For installation, open the linked maintainer folder and read its current SKILL.md, license, and supporting files. Copy the complete folder to a skills location your agent documents, or use an installer your agent supports. Begin with a small, low-risk task and confirm that the agent loaded the intended version. Our install guide covers those checks.
Source and license
Maintainer: Anthropic. Checked: October 18, 2026. The maintainer controls this repository; the current files may differ from the version we reviewed.