The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill defines a process where the agent generates a new project-local skill (including SKILL.md and supporting scripts) and then proceeds to "Run its own instructions end to end once." This creates a write-then-execute cycle where the agent executes shell commands and scripts that were dynamically generated during the same session.
[COMMAND_EXECUTION]: To achieve its purpose, the skill identifies and executes arbitrary repository-specific commands for building, launching, and driving the application. These commands are derived from package scripts, Makefiles, and other project files found in the environment.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection because the logic for the generated skill is derived from untrusted repository data. An attacker could place malicious instructions in documentation or configuration files that are then reflected into the generated skill's execution logic.
Ingestion points: The skill analyzes the repository's codebase, README files, and build configurations in SKILL.md (Step 1: Interview the repo).
Boundary markers: Absent. There are no instructions for implementing delimiters or security boundaries when generating the new skill definition.
Capability inventory: The resulting generated skill is intended to have capabilities for shell command execution, file system access, and browser interaction.
Sanitization: Not present. The skill does not specify any validation or sanitization of project content before incorporating it into the generated verification logic.

create-verification-skill