The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external PRD files and existing local project codebases to influence the generation of a new project's configuration, source code, and build scripts.
Ingestion points: The skill reads docs/prd.md and searches for sibling project files using SoloGraph MCP tools like mcp__solograph__project_code_search and mcp__solograph__codegraph_explain to identify patterns to adopt.
Boundary markers: No explicit boundary markers or "ignore embedded instructions" warnings are mentioned for the ingested PRD or sibling project data.
Capability inventory: The skill has the capability to execute shell commands via Bash (for project setup and verification), use the gh CLI for repository management, and write files using the Write tool.
Sanitization: There is no mention of sanitization, escaping, or validation of the external content before it is interpolated into the generated project's Makefiles or source code.
[DYNAMIC_EXECUTION]: The skill generates new agent instructions and workflow definitions that govern how AI agents will operate within the newly created project.
Evidence: Step 8 generates a project-specific workflow skill in .claude/skills/dev/SKILL.md, and Step 10 creates a CLAUDE.md file containing instructions for agent behavior, such as "Agent behavior defaults" and "Quality Gate" logic.
[COMMAND_EXECUTION]: The skill executes shell commands to automate the project setup and integrate with external version control services.
Evidence: It uses Bash for file system operations and git for repository initialization. It also utilizes the gh CLI tool to create private GitHub repositories and push the initial project state.
[EXTERNAL_DOWNLOADS]: The skill triggers the download of project dependencies as part of its verification and setup process.
Evidence: Step 17 (Verification) instructs the agent to run commands such as pnpm install or uv sync, which fetch software packages from the official npm and PyPI registries. These downloads from well-known services are documented neutrally and do not escalate the verdict.

solo-scaffold