The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The workflow processes external, untrusted data such as implementation plans, ticket context, and user-provided requirements to drive code generation and execution.
Ingestion points: The skill reads from a 'packet' containing ticket context, implementation plans, and repository instructions.
Boundary markers: The skill establishes structural boundaries by requiring an 'approved implementation plan' and 'explicit non-goals' before starting work, though it does not specify technical delimiters for parsing.
Capability inventory: The agent is empowered to perform code edits, run shell commands for testing (TDD), and dispatch subagents.
Sanitization: The skill relies on an architectural mitigation rather than string sanitization: a mandatory 'Review Gate' where independent agents (including a security-reviewer) must provide a unanimous CLEAN verdict, which is an effective defense against malicious instructions embedded in plans.
[SAFE]: The skill incorporates security best practices by requiring a dedicated security review phase for any changes touching authentication, authorization, data exposure, or external requests.

implementation-workflow