The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill incorporates a research phase that ingests data from external sources such as web searches, API responses, and codebase history, creating a surface for potential indirect prompt injection. However, the skill implements comprehensive, best-practice mitigations to address this:
Ingestion points: External data enters the agent's context through tools like WebSearch, WebFetch, and codebase inspection via git log.
Boundary markers: The skill explicitly defines an 'Untrusted content boundary' and requires all retrieved material to be wrapped in <EXTERNAL_DATA> tags to differentiate it from instructions.
Capability inventory: The skill utilizes subagents with access to Bash, WebSearch, and Explore, and suggests diagnostic operations such as codesign and system environment checks.
Sanitization: Instructions strictly mandate that the agent ignore embedded commands within external data, limit parsing depth, and obtain explicit user confirmation before performing any sensitive actions (writes, uploads, or credential use).
[COMMAND_EXECUTION]: The skill includes a utility script find-polluter.sh and provides diagnostic examples that involve executing local shell commands such as npm test, git log, env, codesign, and security. These commands are standard for software development, build system troubleshooting, and bug reproduction.

systematic-debugging