The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of specification files (specs/*.md) and implementation plans. These files drive the autonomous loop's logic, including subagent delegation and code implementation, creating a surface for injection to influence the agent's autonomous behavior.
Ingestion points: The skill reads from specs/*.md, IMPLEMENTATION_PLAN.md, and existing source code files in src/* to build context for each iteration.
Boundary markers: There are no instructions for isolating or delimiting external content to prevent the agent from following instructions embedded within those files.
Capability inventory: The skill has high-impact capabilities, including the Agent tool for massive parallel delegation (up to 500 subagents) and the Bash tool for file modifications and command execution.
Sanitization: No sanitization, escaping, or validation of the ingested markdown content is performed before it is used to shape the agent's plan.
[COMMAND_EXECUTION]: The skill explicitly directs the agent to execute shell commands using the Bash tool for various tasks including testing and environment exploration.
Evidence: The instructions mandate selecting tasks to "Implement" and "Test," specifically recommending the use of the Bash tool with run_in_background=true for non-blocking validation.
[DYNAMIC_EXECUTION]: The skill's primary function is to generate and execute code at runtime to achieve project specifications.
Evidence: In the "BUILDING MODE," the agent is instructed to "Write complete, production-quality code" and "Run tests immediately after implementation," which involves the runtime generation and execution of scripts based on the logic determined during the "PLANNING" phase.

autonomous-loop