The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data from project files and external research, representing a standard surface for indirect prompt injection.
Ingestion points: The research phase in SKILL.md involves reading repository configurations, existing tests, and external documentation.
Boundary markers: Explicit instructions in SKILL.md specify that "Findings never auto-enter the plan" and emphasize that subagent output must be treated as evidence rather than automatic acceptance.
Capability inventory: The skill utilizes file writing (specifically targeting the adrs/ directory) and subagent delegation for research and review tasks.
Sanitization: The workflow requires a manual synthesis step by the main agent and a mandatory adversarial review process to disposition findings before they are finalized.

create-plan