The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill includes a "Self-Evolving Skill" directive instructing the agent to immediately modify the SKILL.md file if it encounters issues or requires workarounds during execution. This behavior creates a risk where untrusted data could influence the skill's permanent instructions.
Ingestion points: External data is ingested from research reports generated by Gemini, which include information scraped from various websites across the internet.
Boundary markers: There are no explicit delimiters or warnings provided to the agent to prevent it from following instructions that might be embedded within the research results.
Capability inventory: The skill has access to Write tools (allowing it to modify its own files) and Bash (allowing it to execute commands).
Sanitization: No sanitization or filtering is applied to the research output before the agent is invited to use it as a basis for "evolving" the skill.
[COMMAND_EXECUTION]: The skill provides instructions for the agent to execute TypeScript files via npx tsx and requires the user to launch a Chrome browser with the --remote-debugging-port=9222 flag. This allows the agent to control a browser instance that is authenticated with the user's Google account.
[EXTERNAL_DOWNLOADS]: The skill relies on npx tsx for execution, which may download the tsx package at runtime if not already cached. It also lists playwright-core as a prerequisite for its automation scripts.

gemini-deep-research