The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local script (node .specforge/core/scripts/instructions.mjs) which is part of its internal framework to manage workflow state. This is a benign use of project-local tooling.
[EXTERNAL_DOWNLOADS]: The agent is instructed to research technical details (versions, pricing, documentation) from well-known and official sources like GitHub, npmjs, and vendor websites. These activities are restricted to factual information gathering and are not used to execute untrusted remote code.
[PROMPT_INJECTION]: While the skill processes user-provided intake documents that could contain untrusted data, it implements a 'Socratic Dialogue Protocol' requiring the agent to ask single questions and await user confirmation. This human-in-the-loop architecture maintains control and prevents the agent from autonomously following instructions potentially embedded in the documents.

sf-brainstorm