The Agent Skills Directory

[PROMPT_INJECTION]: The skill acts as an orchestrator that processes external implementation plans and delegates tasks to subagents. This architecture creates a surface for indirect prompt injection where a malicious plan file could contain instructions designed to influence subagent behavior or bypass intended constraints.
Ingestion points: The PLAN_FILE is processed via scripts/task-brief to generate requirements for implementer subagents.
Boundary markers: Requirements are isolated into brief files and subagents are instructed to treat these files as the source of truth, but no explicit sanitization is performed on the plan text.
Capability inventory: Subagents have capabilities to modify the local filesystem, execute shell commands for testing purposes, and perform git commits.
Sanitization: The skill relies on a multi-stage review process (using task-reviewer-prompt.md and spec-reviewer-prompt.md) where separate subagents verify that implementation work matches the specification and adheres to quality standards.
[COMMAND_EXECUTION]: The skill utilizes several local bash scripts (scripts/sdd-workspace, scripts/review-package, scripts/task-brief) to manage plan-specific workspaces and extract data from git and plan files. These scripts leverage standard system utilities like git, awk, and mkdir to manage artifacts and implementation state.

subagent-driven-development