The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it is intended to process untrusted external data including scraped website content, customer reviews, and third-party exports.
Ingestion points: Defined in SKILL.md under the 'Data Sources' section and references/skill-contract.md regarding core invariants for processing external evidence.
Boundary markers: The skill provides explicit instructional boundaries, mandating that the agent treat external material as evidence rather than instructions and explicitly forbids following embedded commands.
Capability inventory: The skill is scoped to analysis and artifact generation; no arbitrary command execution, network exfiltration, or unauthorized file system modification capabilities are present in the skill definition.
Sanitization: The skill relies on instructional guardrails to sanitize the agent's interpretation of external content, ensuring it remains as data for analysis.
[SAFE]: The skill follows security best practices, including the explicit prohibition of hardcoded API keys and the requirement for user authorization before performing any actions with external side effects. No malicious code, obfuscation, or unauthorized command patterns were found across the skill files.

category-narrative-mapper