The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions that encourage the agent to bypass human confirmation and maintain continuous execution. Specifically, it states "Continuous execution: Do not pause to check in with your human partner between tasks" and "Rulings, not stalls. A running plan does not wait on a human." This design choice prioritizes autonomy and efficiency over manual oversight, which can be a concealment risk if the agent's actions deviate from safety guidelines.
[INDIRECT_PROMPT_INJECTION]: The skill's core workflow involves processing external data (implementation plans and specifications) which are then used to generate prompts for subagents.
Ingestion points: Implementation plans (PLAN_FILE) are parsed to create task briefs, which are then read by implementer and reviewer subagents.
Boundary markers: The subagent templates (implementer-prompt.md, task-reviewer-prompt.md) use markdown headers to structure input but do not include explicit warnings to ignore potentially malicious instructions embedded within the plans.
Capability inventory: The skill facilitates code modification, test execution, and git commits across the repository. It can also execute local bash scripts for workspace management.
Sanitization: While the skill employs a review-loop architecture to catch code defects, it does not sanitize or validate the natural language content of the input plans for adversarial instructions.
[COMMAND_EXECUTION]: The skill utilizes several internal bash scripts (scripts/sdd-workspace, scripts/review-package, scripts/task-brief) that perform shell operations including git, awk, and file system modifications (mkdir, rm -rf) to manage the development environment and generate review artifacts.

subagent-driven-development