The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill implements self-modifying instructions. According to swan.md, the agent uses the swan-update-skill tool to rewrite the 'Setup state' paragraph in SKILL.md upon completing the configuration process. This allows the skill to dynamically update its own logic to persist state.
[DYNAMIC_EXECUTION]: The skill uses the swan-execute-code tool to aggregate deal data and parse content from user-uploaded files, including PDFs using the pypdf library. This involves executing logic on data provided by external, untrusted sources.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from multiple sources, creating an attack surface for indirect prompt injection.
Ingestion points: Content is retrieved from external URLs via swan-fetch-scraped-url and from user-provided documents (Path C).
Boundary markers: The instructions include mandatory human-in-the-loop steps ('Show back, confirm, save') and 'NEVER auto-save' rules to prevent unintended execution of instructions found in external data.
Capability inventory: The skill has access to code execution (swan-execute-code) and file modification (swan-update-skill) tools.
Sanitization: There are no specified sanitization or validation routines for the data fetched from external sources.

icp