The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data (such as technical specifications, design documents, and RFCs) which are then analyzed by a reviewer LLM.
Ingestion points: Untrusted data enters via the spec-path and --context arguments, which are passed to the reviewer prompt defined in references/peer-review-prompt.md.
Boundary markers: The skill employs a 'SCOPED-WRITE CONTRACT' in the prompt to define output limits and instructions for the reviewer LLM, though this does not mitigate risks from adversarial content within the processed documents.
Capability inventory: The skill includes the capability to execute local bash scripts (scripts/validate-findings.sh) and write reviewer findings to the local filesystem.
Sanitization: The instructions do not specify any validation or sanitization for the content of the specifications being reviewed.
[COMMAND_EXECUTION]: The skill logic involves the construction and execution of shell commands that include variables derived from user input.
Evidence: In SKILL.md (Step 5), the command bash <skill-dir>/scripts/validate-findings.sh --kind techspec --round <N> --path <out>/peer-review-findings-roundN.md uses user-provided variables <N> and <out>. This creates a potential command injection surface if the underlying agent platform does not sanitize or safely escape these parameters before execution.

spec-peer-review