The Agent Skills Directory

Indirect Prompt Injection Guardrails: The skill recognizes that catalog metadata (descriptions, READMEs, glossary terms) and source documents (S3-hosted markdown files) are untrusted data that could contain malicious instructions. It establishes a mandatory rule (Rule 3) and a dedicated safety reference to ensure the agent treats this content strictly as data, avoids acting on embedded commands, and performs content extraction rather than verbatim copies.
Data Integrity and Recovery Patterns: Because certain AWS catalog APIs (such as Glue UpdateTable or DataZone CreateAssetRevision) replace existing records rather than merging them, the skill enforces a "snapshot, write, verify, restore" pattern. This practice involves backing up current metadata to a local file before modification, ensuring that accidental clearing of fields can be detected and reversed immediately.
Secure Dynamic Execution: The skill generates Bash and jq scripts to facilitate bulk operations and metadata updates. These scripts employ secure interpolation techniques, such as using jq --arg or --argjson to pass data safely, which mitigates risks of command or filter injection during the execution of generated logic.
Sensitive Data Protections: Clear instructions are provided to prevent the accidental disclosure of secrets, tokens, PII, or internal account identifiers within the catalog metadata. The skill also recommends enabling AWS CloudTrail with log file validation and encryption to maintain an auditable record of all catalog mutations.
Contextual Environment Verification: The skill includes mandatory platform detection probes (e.g., sts get-caller-identity, datazone list-domains) to ensure the agent identifies the correct AWS environment and principal before proposing or executing changes, reducing the risk of operations in an unintended account or region.

managing-data-catalog