The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied objectives and requirement refinements which are then directly interpolated into a high-level instruction template for a 'reviewer subagent'. This creates a vulnerability surface where a malicious user prompt could attempt to override the reviewer's instructions or behavior.
Ingestion points: The [original prompt] and [summary of results of grilling session] variables are injected into the reviewer subagent prompt in SKILL.md.
Boundary markers: None detected. There are no delimiters or 'ignore embedded instructions' warnings surrounding the user-controlled inputs.
Capability inventory: The skill uses Bash for command execution, agent-browser for web automation, git for branch management, and github-cli for pull request creation.
Sanitization: No evidence of sanitization, escaping, or validation of the input data before prompt interpolation.
[DYNAMIC_EXECUTION]: As part of the review process, the subagent is instructed to 'Build and run the system with Bash' and 'invoke it with realistic inputs'. This involves the automated execution of code that was dynamically generated by the agent based on user prompts.
[COMMAND_EXECUTION]: The skill utilizes shell commands (git, gh, bash) to perform its primary functions, including repository management and pull request delivery. This is a standard capability for developer-oriented skills but represents a powerful execution environment.

freehand