The Agent Skills Directory

[PROMPT_INJECTION]: Indirect Prompt Injection Surface. The skill is designed to ingest and act upon data from external markdown files (docs/mvp-first-draft.md, docs/world-lore.md, etc.), which could contain malicious instructions intended to influence the agent's code generation logic.
Ingestion points: docs/mvp-first-draft.md, docs/build-registry.md, docs/world-lore.md, and docs/quest-registry.md are read during Phase 0 and Phase 4 to determine build logic and narrative coherence.
Boundary markers: The skill instructions do not explicitly define boundary markers or instruct the agent to ignore Natural Language instructions found within the project documentation files.
Capability inventory: The skill has the capability to write and modify files within the src/ and docs/ directories and suggests shell commands (bun vitest, bun run dev) for the user to execute.
Sanitization: No explicit sanitization or validation of the content of the markdown files is performed before the data is used to generate executable code.
[COMMAND_EXECUTION]: Shell Command Suggestions. The skill instructs the agent to provide the user with shell commands for running and testing the generated components.
Evidence: Phase 6 specifies the output of commands such as bun run dev and bun vitest run src/[component-name]/. While these are suggested to the user rather than executed autonomously, they represent a path for potentially malicious code execution if the generated components were manipulated via injection.

game-build