The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill researches improvements via web search and incorporates the results into new skill variations. This allows untrusted external data to influence the instructions generated and executed by the agent.
Ingestion points: Web search results (Step 2) and target SKILL.md files (Step 1).
Boundary markers: None explicitly defined to separate researched content from the generation prompt.
Capability inventory: Filesystem writes (SKILL.md), git push, gh pr create, and curl operations.
Sanitization: No sanitization or validation of external content is performed before interpolation into the new skill code.
[DYNAMIC_EXECUTION]: The skill dynamically generates and overwrites SKILL.md files, which function as the executable code for the agent platform. The generation logic relies on external, untrusted research data.
[COMMAND_EXECUTION]: The skill uses powerful CLI tools like git and gh (GitHub CLI) to commit and push changes and create Pull Requests. This automates the persistence and deployment of dynamically generated instructions.

autoresearch