The Agent Skills Directory

[COMMAND_EXECUTION]: Every workflow file (Dream.md, FullCycle.md, Mate.md, QuickCycle.md, Steal.md, Test.md) includes instructions to execute curl commands. These commands perform POST requests to http://localhost:31337/notify to provide local status notifications.
[COMMAND_EXECUTION]: The skill maintains an execution log by appending JSON-formatted entries to ~/.claude/PAI/MEMORY/SKILLS/execution.jsonl using the echo command and shell redirection at the conclusion of each workflow.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its core functionality of ingesting and iteratively processing untrusted external data.
Ingestion points: Untrusted content enters the system during the CONSUME and STEAL phases, which invoke the Research skill to gather information from external domains and patterns (Workflows/FullCycle.md, Workflows/Steal.md).
Boundary markers: The instructions do not define the use of delimiters or 'ignore' instructions when presenting researched data to subsequent agent phases (DREAM, MATE, etc.).
Capability inventory: The system can perform network requests (curl), file writes (echo), and invoke various other skills and tools.
Sanitization: There is no evidence of sanitization, validation, or escaping of the external content before it is interpolated into agent prompts throughout the evolutionary loop.

Ideate