The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill maintains an attack surface for indirect prompt injection because its primary function is to ingest and analyze untrusted project data (source code, schemas, and configurations).
Ingestion points: Injects repository content including schemas, routes, SDKs, and versioned configurations from the project being audited (specified in SKILL.md).
Boundary markers: Implements explicit 'checkpoint' and 'mandatory pause' markers in the workflow to require human/legal validation before sensitive actions.
Capability inventory: Can read files, write to the .lgpd/ directory, and programmatically invoke 18 specialized sub-skills via agent tools.
Sanitization: Relies on human-in-the-loop checkpoints and explicit confirmation steps rather than automated sanitization of ingested code content.
[COMMAND_EXECUTION]: The orchestrator triggers subordinate skills using the platform's native tool-calling mechanisms (e.g., skill or activate_skill) to navigate the privacy audit pipeline as described in references/skill-activation.md.
[EXTERNAL_DOWNLOADS]: Normative references in references/normative-reference.md point to official Brazilian government websites (planalto.gov.br and gov.br). These are trusted domains for legal and regulatory documentation and do not pose a security threat.

lgpd-audit