The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest HTML content from external URLs or the Stitch API and process it to generate new code. This workflow is vulnerable to indirect prompt injection if the source content contains hidden instructions meant to manipulate the agent during the conversion process.
Ingestion points: External content is downloaded to temp/source.html using the scripts/fetch-stitch.sh script.
Boundary markers: The instructions do not specify any delimiters or safety prompts to ensure the agent ignores instructions potentially embedded in the source HTML.
Capability inventory: The skill has access to shell execution (Bash), the file system (Read, Write), and specialized tools (stitch*:*).
Sanitization: No sanitization or filtering logic is defined for the content before it is processed by the agent.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to execute a local utility script, scripts/fetch-stitch.sh, which takes a URL as an argument to perform downloads. While the script uses quoted variables, the execution of shell commands based on external inputs is a notable behavior.
[EXTERNAL_DOWNLOADS]: The skill facilitates the retrieval of content from external sources, including Google Cloud Storage (GCS) and user-specified URLs, to the local environment for processing.

stitch-html-components