The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes dynamic context injection (the ! syntax) to automatically execute shell discovery commands such as cat, find, head, and xargs when the skill is loaded. These commands are used to identify the project's framework, manifest files, and shared configuration patterns.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by reading and processing untrusted data from local project files like README.md, package.json, and Fastlane metadata.
Ingestion points: File contents from the project root and subdirectories are read into the agent's context during discovery and extraction steps.
Boundary markers: There are no explicit delimiters or instructions telling the agent to treat the ingested project data as untrusted or to ignore instructions embedded within those files.
Capability inventory: The skill is capable of writing to the local file system (e.g., creating .appshot-context.json) and provides guidance for running project build scripts.
Sanitization: No validation or filtering is applied to the ingested text before it is used for feature extraction and category inference.
[DATA_EXFILTRATION]: The skill instructions direct the agent to check .env and .env.local files to extract public configuration values like NEXT_PUBLIC_APP_URL. Although targeted at public-facing variables, environment files are sensitive locations that often contain private credentials, making automated access a potential exposure risk.

appshot-core