The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes a persistent indirect prompt injection surface by creating project blueprint files (AGENTS.md and handoff.md) which the agent is explicitly instructed to read at the beginning of every session.
Ingestion points: Data is ingested from user-provided project descriptions/goals and from the results of a local directory scan during initialization (SKILL.md).
Boundary markers: No boundary markers, XML tags, or "ignore embedded instructions" delimiters are specified for the content generated in the blueprint files.
Capability inventory: The skill possesses the capability to write to the local file system and execute shell commands (gh, git, hostname) to populate these files.
Sanitization: The skill does not implement sanitization, escaping, or validation for the data interpolated into the generated markdown files, allowing potentially malicious filenames or project metadata to influence future agent sessions.
[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands to gather environment metadata and verify tool status.
Evidence: The instructions include calls to gh auth status to verify GitHub CLI login, and platform-specific commands such as $env:COMPUTERNAME (PowerShell) or hostname (Mac/Linux) to identify the machine (SKILL.md).

project-init