The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources such as project fixtures and databases to perform its primary validation function.
Ingestion points: Reads content from files in directories like testdata/, fixtures/, and seeds/, as well as from configured database connections.
Boundary markers: The instructions do not define specific delimiters or "ignore" directives to prevent the agent from accidentally following instructions if they are found within text-based data fields during validation.
Capability inventory: The agent has capabilities to read local files, execute database SELECT and EXPLAIN queries, and write report files to the local disk.
Sanitization: While the skill enforces PII masking (redacting emails, phones, names) in its output reports, it does not specify sanitization or escaping of the input data to neutralize potential embedded prompts.
[CREDENTIALS_UNSAFE]: The skill is instructed to access sensitive environment configuration files to resolve database connection strings.
Evidence: Phase 1.1 explicitly directs the agent to look for a DATABASE_URL within .env.test or .env.local files to establish a data source connection.
Context: This access is limited to non-production targets defined within the repository and is a requirement for the skill's intended purpose of data validation.

data-validation