The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow where raw user input from a Telegram command (/meta) is passed directly into an agent session with project-wide modification capabilities.
Ingestion points: User-supplied text provided after the /meta command in a Telegram interface.
Boundary markers: The documentation lacks any mention of delimiters or instructions for the agent to disregard embedded commands within the user prompt.
Capability inventory: The integration with the Claude Code Agent SDK grants the ability to read and modify local files, as well as execute commands within the ~/ai_projects/telegram_agent directory.
Sanitization: No validation, sanitization, or filtering of the user prompt is specified before it is executed by the service.
[DYNAMIC_EXECUTION]: The skill is designed to enable self-modifying code behavior, which is a significant security and stability risk.
Evidence: The skill's description explicitly states its purpose is to 'fix bugs, add features, or modify the bot code' of the project it is currently running in. This allows for runtime modification of the agent's own logic via natural language commands, which could be exploited to introduce backdoors or disable safety constraints.

meta