The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill documentation indicates that execution guidance may include the command npx skills add -y -g <source> for installing external skills. The inclusion of the -y flag automatically accepts installation prompts, and since the source is not restricted, an agent following these instructions could download and execute malicious code from an untrusted remote repository.
[COMMAND_EXECUTION]: The skill relies on a suite of vendor-specific CLI tools (joelclaw) to perform core functions such as planning, dispatching, and running workloads. These tools interact with the local file system (e.g., --write-plan ~/.joelclaw/workloads/) and potentially the network (e.g., --send-mail).
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and act upon user-provided strings described as 'intent' and 'steering inputs'.
Ingestion points: Found in SKILL.md and references/common-workloads.md, specifically the <intent> argument for the planning command and input for Acceptance: clauses.
Boundary markers: No specific boundary markers or 'ignore' instructions are provided to prevent the agent from following malicious commands embedded within the user's intent.
Capability inventory: The skill has broad capabilities including file writing, command execution via joelclaw, and remote code installation via npx.
Sanitization: The skill lacks any description of sanitization or validation for the user-provided steering inputs before they are used to generate execution plans.

agent-workloads