The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external sources and user input, which is then processed by specialized agents in a multi-step pipeline. Specifically, the Phase 1 and Phase 3 agents are instructed to fetch content from resource_url or github_url found in research artifacts. This content is used to inform the architecture spec and task plan, creating an attack surface where malicious instructions in external documents could influence the agent's behavior.
Ingestion points: The workflow reads from $ARGUMENTS, local artifacts via artifact_read, and external URLs via mcp__Ref__ref_read_url or WebFetch (Phase 1 and 3).
Boundary markers: The skill uses delimiters such as <feature_request>, <prior_artifacts>, and <quality_vigilance> to separate injected content from instructions.
Capability inventory: The skill can write to the filesystem (via artifact registration), create executable plans (sam_plan), perform network requests (WebFetch), and search files (Glob).
Sanitization: The skill relies on the <quality_vigilance> block, which instructs agents to verify claims against primary sources and note concerns, rather than programmatic sanitization or escaping.
[COMMAND_EXECUTION]: The workflow relies on executing shell-based commands through <sam_cli/> and various MCP tools for project management tasks, such as artifact list, artifact read, backlog view, and backlog update. These tools are used to manage the project state and persist design artifacts.

add-new-feature