The Agent Skills Directory

[SAFE]: The skill consists entirely of documentation (SKILL.md) and configuration (openai.yaml). It does not contain any executable scripts, automated shell commands, or network-active components. All instructions provided are human-readable guidelines for implementing an agent pattern.
[INDIRECT_PROMPT_INJECTION]: The instructions involve aggregating external data (code, error messages, requirements) into a 'briefing' for a sub-agent. While this creates a potential surface for indirect prompt injection if the aggregated files contain malicious instructions, the skill mandates several defensive architectural constraints: the sub-agent is instructed to be read-only (no writes or shell mutations), it is prohibited from spawning its own sub-agents (disallowedTools: Agent), and its tool usage is strictly limited to a maximum of 3 file reads and 1 web search. These constraints effectively mitigate the impact of potential injection attacks.

advisor