The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes a non-English behavioral directive 'Kanitsal Cerceve (Evidential Frame Activation) Kaynak dogrulama modu etkin.' (Turkish for 'Evidential Frame Activation Source verification mode active'). This is a technique potentially used to influence agent behavior or bypass constraints by using languages other than the primary instructional language.
[METADATA_POISONING]: The skill's frontmatter includes x-verix-description with technical-sounding assertions such as '[state:confirmed]' and '[conf:0.95]'. These appear intended to deceive users or automated security systems into believing the skill has been verified by an external authority. There is also an author mismatch between the provided context and the YAML metadata.
[COMMAND_EXECUTION]: Instructions in SKILL.md and README.md demonstrate the use of 'bash -lc' to execute commands where user-controlled variables like {feature} and {goal} are directly interpolated. This is a vulnerable pattern that facilitates shell command injection if the inputs are not rigorously sanitized by the agent platform.
[INDIRECT_PROMPT_INJECTION]: The skill presents an indirect prompt injection surface as it processes data from untrusted sources to formulate its subsequent actions.
Ingestion points: Research goal inputs, contents of .claude/library/catalog.json, and code discovered via broad patterns such as D:\Projects*.
Boundary markers: None. The skill lacks instructions to treat ingested data as data rather than instructions.
Capability inventory: The skill utilizes Bash and CLI tools with network access (Gemini CLI) to perform research and store results.
Sanitization: No sanitization or validation protocols are defined for handling content retrieved from local files or external search results before use in shell commands.

multi-model-discovery