The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a persistent 'manager' role that takes over the agent's behavior for the entire session. It also presents an indirect prompt injection surface by delegating work to sub-agents using templates that interpolate untrusted data.
Ingestion points: User task descriptions and sub-agent execution reports are ingested as context for new agent dispatches.
Boundary markers: The templates use plain text headers (e.g., 'Original requirements:') but lack formal delimiters or instructions to ignore embedded commands.
Capability inventory: The skill can spawn background processes (agents-mcp), execute shell commands, and interact with the GitHub API.
Sanitization: No explicit sanitization is performed on user or sub-agent content before interpolation.
[EXTERNAL_DOWNLOADS]: The skill instructions (README.md and AGENTS.md) direct the installation of several external tools, including agents-mcp, compose-agentsmd, @metyatech/task-tracker, and @metyatech/thread-inbox. These are hosted on NPM or GitHub and originate from the author's namespace.
[COMMAND_EXECUTION]: The skill frequently executes CLI tools such as git, gh, npm, and npx for orchestration. It also provides specific PowerShell scripts to bypass local policy restrictions on file deletion.
[DATA_EXFILTRATION]: The skill is designed to automate repository management, including committing, pushing to remote repositories, and creating GitHub releases, which involves sending project code and metadata to external servers.

manager