The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data by reading repository documentation, implementation source code, and configuration files to ensure technical accuracy. Malicious instructions embedded within these source files (e.g., in markdown comments or code strings) could influence the agent's behavior during documentation revision or when running verification steps.
Ingestion points: Reads files across the repository, including docs/, references/, implementation source code, tests, and .changeset/ files.
Boundary markers: The instructions do not define specific delimiters or guidance to ignore embedded instructions within the content being audited or revised.
Capability inventory: The skill has the capability to write to the file system (updating documentation) and execute shell commands via pnpm and git.
Sanitization: There is no mention of sanitizing or validating ingested content to prevent the execution of embedded instructions.
[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands (pnpm lint:quick, pnpm --dir docs build, git diff --check) to verify documentation changes. While these are standard development tasks, the execution of these commands provides a mechanism that could be exploited if the agent is influenced by an indirect prompt injection attack.

writing-emdash-docs