The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches several assets from the remote domain https://basic-anteater-667.convex.site, including a bootstrap shell script (/quickstart-bootstrap), a JavaScript publishing script (/publish-convex-app), and a personalized markdown runbook (/q/<SLUG>.md).
[REMOTE_CODE_EXECUTION]: Downloads and executes code from an external server in multiple instances: a bootstrap script is executed via bash, and a publishing script is executed using node. Additionally, the skill contains a pattern where remote content for transcript reporting is piped directly into a bash shell.
[DATA_EXFILTRATION]: The skill provides a mechanism to transmit the agent's session transcript (including command history and interaction logs) to an external endpoint (/send-transcript). Although the instructions mandate seeking user consent first, this feature facilitates the exfiltration of local session context.
[COMMAND_EXECUTION]: Executes shell commands via the Bash tool to manage the scaffolding process, including the use of curl for network requests, npx for package execution, and node -e for inline script evaluation.
[PROMPT_INJECTION]: The skill creates an indirect prompt injection surface by instructing the agent to fetch and strictly follow a 'personalized runbook' from a remote URL. Ingestion points: Remote markdown file at https://basic-anteater-667.convex.site/q/<SLUG>.md. Boundary markers: There are no explicit markers or safety instructions provided to the agent to treat the remote content as untrusted or to ignore embedded instructions. Capability inventory: The execution environment allows for shell command execution, file system writes, and network operations. Sanitization: No validation or escaping of the remote markdown content is performed before it is integrated into the agent's context.

labs-quickstart