The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes an architecture where user answers from onboarding.json are stored in agent_data/USER.md and subsequently interpolated into system prompts via the {{include}} directive. This provides a clear surface for indirect prompt injection if the inputs are not properly sanitized.
Ingestion points: User-provided answers are written to agent_data/USER.md within the sandbox.
Boundary markers: The documentation does not specify the use of delimiters or 'ignore' instructions when including the user data into prompts.
Capability inventory: The described agent environment has bash/read/write access to the sandbox, can perform network requests for visual content, and can query a SQLite activity log.
Sanitization: There is no mention of sanitization or validation of the onboarding answers before they are processed by the agent's prompts.
[EXTERNAL_DOWNLOADS]: The documentation instructs developers to install a CLI tool directly from a specific GitHub branch (github:sylm54/train-me#framework-cli) using package managers like npm, pnpm, or bun.
[COMMAND_EXECUTION]: The skill provides instructions for executing the framework CLI using bunx tm-framework, which runs commands from the external package to perform linting and packaging tasks.
[DATA_EXPOSURE]: The TTS XML tag system documentation describes a <visual> tag that performs network requests to redgifs.com to fetch, cache, and display content based on user-configurable niche and tag metadata.

framework-dev