The Agent Skills Directory

[SAFE]: The skill mandates robust data protection measures, requiring the redaction of API keys, authorization headers, prompts, documents, and customer identifiers from all telemetry and outputs.
[SAFE]: External references and package recommendations (cohere-ai@8.1.0, cohere==7.1.1) target official Cohere repositories (github.com/cohere-ai) and domains (docs.cohere.com, cohere.com), which are recognized as well-known service resources.
[SAFE]: The skill provides explicit instructions to manage credentials securely, prohibiting the persistence or printing of API keys and recommending injection from approved secret managers.
[INDIRECT_PROMPT_INJECTION]: A vulnerability surface for indirect prompt injection exists as the skill is designed to process external documentation and repository contents.
Ingestion points: Files are read from the repository using Read, Glob, and Grep, and external documentation is retrieved via WebFetch (SKILL.md, references/official-docs.md).
Boundary markers: Not explicitly defined for delimiting untrusted content within prompts.
Capability inventory: The skill uses Read, Glob, Grep, WebFetch, Write, and Edit tools.
Sanitization: The skill contains extensive instructions for redacting sensitive content, allowlisting telemetry fields, and requiring manual approval for sensitive side effects.

cohere-observability