The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill instructs the agent to calculate weighted totals for its rubric scores by executing a Python script via the command line. The command template provided in SKILL.md (python3 -c "w=(3,3,2,1,1);r={'Idea A':(8,7,7,9,7),...}") interpolates idea names directly into a Python dictionary literal. Because these idea names are derived from user-supplied briefs and previous agent outputs, an adversary could craft a brief that causes the agent to generate an idea name containing escaping characters (like single quotes) and arbitrary Python code. This would result in the execution of that code within the agent's shell environment.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves ingesting and processing untrusted data from user briefs and rejected previous outputs (Step 1 and Step 2). This information is used to define the 'mode' and 'fence' tiers, which in turn drive the generation of new content and the use of tools such as web search and CLI-based repository searches (gh search). There are no instructions for sanitizing this input or using clear boundary markers to prevent malicious instructions embedded in the brief from influencing the agent's execution flow.
[COMMAND_EXECUTION]: The skill relies on several shell-based tools for its operations, including python3 for mathematical calculations and random word sampling, as well as grep and gh search for validation checks. While these are legitimate utility uses, the parameters for these commands are often constructed using data derived from external or user-influenced sources, which necessitates careful handling to prevent command injection.

creative