The Agent Skills Directory

[COMMAND_EXECUTION]: The skill frequently executes subcommands of the openspec CLI (such as list, status, instructions, and new change) to perform its primary functions.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection by processing external data.
Ingestion points: The agent reads JSON output from openspec status and openspec instructions, and also reads the content of existing 'dependency' files from the local disk.
Boundary markers: The instructions provide a specific guardrail telling the agent to treat context and rules as internal constraints and to not copy them into the output artifacts.
Capability inventory: The agent has the ability to execute openspec commands and write files to paths (resolvedOutputPath) provided by the tool.
Sanitization: There is no mention of sanitizing or validating the content of the dependency files before the agent uses them to generate new artifact content.
[DYNAMIC_EXECUTION]: The skill includes instructions to delegate the creation of artifacts to other skills or commands if specified in the instruction field of the CLI's output. This creates a dynamic control flow where the agent's behavior is dictated by the data returned from an external tool execution.

openspec-continue-change