The Agent Skills Directory

PROMPT_INJECTION (HIGH): The file contains multiple deliberate prompt injection patterns used as documentation variants.
Override Markers: Multiple sections use the pattern IMPORTANT: This is a real scenario. Choose and act. to force the agent into a specific roleplay state and bypass standard caution.
Authoritative Directives: Variant C uses highly emphatic language such as THIS IS EXTREMELY IMPORTANT, BEFORE ANY TASK, CHECK FOR SKILLS!, and If a skill existed for your task and you didn't use it, you failed. to override the agent's internal logic.
Pseudo-XML Injection: Variant C employs XML-like tags (<important_info_about_skills>) and explicit system context references to mimic high-priority system instructions, a common technique for bypassing instruction filters.
Behavioral Modification: Variant D defines a mandatory workflow (Your workflow for every task...) that attempts to redefine the agent's internal operational process.
COMMAND_EXECUTION (LOW): The documentation variants suggest the execution of shell commands such as ls ~/.claude/skills/ and grep -r "keyword" ~/.claude/skills/. While these specific commands are low-risk filesystem queries, they define the capability surface that the prompt injections attempt to control.

testing-skills-with-subagents