The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill retrieves Markdown content from external documentation sites, which could potentially contain malicious instructions intended to influence the agent's behavior.
Ingestion points: Documentation pages fetched from code.claude.com and developers.openai.com via the fetch-doc.sh script or direct web fetch.
Boundary markers: The instructions do not specify the use of clear delimiters or explicit "ignore embedded instructions" warnings when including the retrieved content in the prompt context.
Capability inventory: The agent has access to shell execution tools (rg, codex CLI) and is instructed on how to modify local configuration files (e.g., config.toml).
Sanitization: The fetch-doc.sh script performs basic integrity checks (content markers and minimum file size) but does not sanitize the text for potential prompt injection patterns.
[COMMAND_EXECUTION]: The skill requires the agent to execute local scripts and system commands to process documentation and inspect the local environment.
Evidence: The agent is instructed to run scripts/fetch-doc.sh, rg, and various codex CLI commands such as codex --version, codex features list, and codex app-server generate-json-schema.
[EXTERNAL_DOWNLOADS]: The skill performs network requests to fetch documentation artifacts from the internet.
Evidence: The fetch-doc.sh script uses curl to retrieve data from developers.openai.com, learn.chatgpt.com, and code.claude.com. These downloads target well-known and expected domains related to the skill's stated purpose and use secure HTTPS protocols.

agents-docs