The Agent Skills Directory

[DATA_EXPOSURE]: The skill explicitly requires the agent to 'Strip the specifics' and 'Generalize the pattern'. It mandates replacing real company or contact names with generalized placeholders like 'the target company' or 'the account on the list'. It also forbids hard-coding filter values or segment definitions, instead instructing the agent to load these from organizational memory. This approach significantly reduces the risk of accidental data exposure or hard-coding sensitive information.
[COMMAND_EXECUTION]: The skill references platform-specific tools including swan-execute-code, swan-build-sequence, and swan-create-skill. These tools are used within their intended scope for managing skill creation and execution of internal logic. The skill includes a 'human-in-the-loop' requirement, mandating that the agent MUST confirm the title, category, and scope with the user before invoking the swan-create-skill tool.
[INDIRECT_PROMPT_INJECTION]: As the skill processes conversation history to generate new instructions, it naturally possesses an indirect injection surface. However, the skill contains robust mitigation strategies: it specifically instructs the agent to only skillify 'repeatable' and 'already proven' work, and to 'strip and generalize' the output. These constraints act as filters against the accidental adoption of malicious instructions found in untrusted conversation data.

skillify