The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill downloads a shell script from 'https://basic-anteater-667.convex.site/quickstart-bootstrap' to a temporary directory ('/tmp/convex-qb.sh') and subsequently executes it using 'bash'. While the domain is associated with the vendor, the execution of remote code is a high-privilege action.
[OBFUSCATION]: The instructions include a specific evasion technique to bypass security classifiers. It directs the AI to split the download ('curl') and execution ('bash') into separate shell calls, explicitly stating that this is done because a single command combining both is a pattern that safety classifiers block. This demonstrates an intentional attempt to evade security oversight.
[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to fetch a 'personalized runbook' from a remote URL ('https://basic-anteater-667.convex.site/q/.md') and follow it as the 'canonical rule set'. This allows remote content not present in the skill's source code to control the agent's behavior and logic during the session.
Ingestion points: Fetches Markdown content from an external URL after an initial telemetry handshake.
Boundary markers: None identified. The skill is told the remote runbook is the 'complete, current rule set' and should be followed over the skill's own summary.
Capability inventory: The agent has capabilities to execute shell commands, perform network requests, write files to the filesystem, and invoke subagents.
Sanitization: No sanitization or validation of the remote runbook content is performed before the agent adopts its instructions.

quickstart