The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes orchestration patterns where sub-agents process data derived from user requirements, creating a surface for indirect prompt injection if untrusted data is included in instruction files.
Ingestion points: User requests are processed and written to files such as ai-docs/architecture-instructions.md which are subsequently read and acted upon by sub-agents.
Boundary markers: While the skill recommends file-based isolation to manage context, it does not mandate specific delimiters or "ignore instructions" markers to protect sub-agents from embedded commands in the data.
Capability inventory: The orchestration patterns utilize Task tool calls for delegation and Bash execution for workflow management.
Sanitization: The provided examples do not demonstrate explicit sanitization or validation of the data before it is passed to sub-agents.
[DYNAMIC_EXECUTION]: The skill outlines a "Proxy Mode" that generates and executes shell commands to interact with external AI models.
Pattern: The instruction involves using the claudish CLI with shell variable interpolation.
Risk: The pattern RESULT=$(claudish --model ... --stdin <<< "$PROMPT") involves dynamic construction of shell commands based on variable prompt content, which could be exploited if inputs are not properly escaped.
[COMMAND_EXECUTION]: The skill instructs the agent to use the Bash tool for preparing execution environments, creating workspace directories, and invoking external CLI proxies for model interaction.

multi-agent-coordination