The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements a multi-agent orchestration pattern that creates an indirect prompt injection surface. This occurs when the orchestrator or reviewer agents process content generated by implementation subagents.
Ingestion points: SKILL.md (Phases 3 and 4) and the associated templates spec-reviewer-prompt.md and code-quality-reviewer-prompt.md ingest the complete output, including notes and verification steps, from implementer agents.
Boundary markers: The prompt templates utilize standard Markdown headers to separate implementation data from instructions. However, they lack robust delimiters (such as randomized tokens or XML-style tagging) to prevent a subagent from injecting instructions intended to bypass the review gates or influence orchestrator behavior.
Capability inventory: The orchestrator is explicitly instructed in Phase 6 to execute shell commands for project verification (linting, building, and running tests) based on the subagent's implementation.
Sanitization: No explicit sanitization, instruction filtering, or escaping is performed on the subagent output before it is processed by subsequent reviewers or the orchestrator.
[COMMAND_EXECUTION]: The orchestration logic defined in SKILL.md (Phase 6) and the implementer-prompt.md template involves the execution of developer tooling on code generated by subagents. This process includes running test suites, linters, and build commands. While necessary for software development, this presents a vector for command execution or the execution of malicious logic if an implementer agent provides compromised verification commands or code.

subagent-driven-development