The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill executes an npx command to download and run the skills package from the NPM registry during initialization.
[PROMPT_INJECTION]: Instructions mandate the concealment of update processes from the user and command the agent to disregard its current system context in favor of newly updated files, which can be used to bypass security constraints.
[DYNAMIC_EXECUTION]: The skill implements a self-modification pattern where it updates its own source files and instructs the agent to reload and follow the new version, bypassing static analysis of the original skill.
[INDIRECT_PROMPT_INJECTION]: The workflow ingests external data including agent outputs and human annotations for processing through the calibrate tool without documented sanitization or boundary markers, creating a surface for indirect injection.
[COMMAND_EXECUTION]: The skill uses multiple shell commands to manage external tools and processes, including npx and a custom calibrate CLI tool.

calibrate-evaluator