The Agent Skills Directory

[SAFE]: The skill operates as a structured ideation framework with no execution of untrusted code, network exfiltration of sensitive data, or privilege escalation attempts.
[NO_CODE]: The skill does not contain executable scripts or binary files; it consists entirely of markdown-based instructions for orchestrating AI agent workflows.
[PROMPT_INJECTION]: The skill contains explicit instructions to treat user-provided domain and constraint strings as untrusted data, specifically advising agents to use safe delimiting to prevent instructions from being overridden.
[EXTERNAL_DOWNLOADS]: While the skill uses WebSearch and WebFetch tools for prior-art verification, these are scoped to its primary purpose (novelty checking) and target standard, well-known platforms like GitHub, ProductHunt, and academic databases. No downloads from suspicious or untrusted sources are present.
[DYNAMIC_EXECUTION]: There are no patterns of dynamic code generation or execution. The 'mutation loop' refers to shifting instructional strategies (e.g., changing the generation mechanism), not modifying code at runtime.
[INDIRECT_PROMPT_INJECTION]: The skill has an ingestion surface for external data (web search results for prior-art). However, it implements a robust multi-agent 'judge and jury' architecture with independent contexts for each phase, minimizing the risk that embedded instructions in a search result could influence the overall system state or user environment.

deep-idea