The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to treat generated prototypes as primary sources (一手资料) for future context and decisions. This creates a surface for indirect prompt injection where instructions hidden within prototype code comments or UI elements could influence the agent's future behavior during implementation phases. Ingestion points: Prototypes saved to once-off branches or project directories (SKILL.md). Boundary markers: Absent; the skill does not specify delimiters or warnings for the agent when reading these files as source material. Capability inventory: The skill possesses file write capabilities and shell command execution permissions (SKILL.md, UI.md). Sanitization: No sanitization is specified for content extracted from prototypes during the saving results phase.
[DATA_EXFILTRATION]: The instructions for Logic Prototypes (LOGIC.md) involve creating self-contained, standalone HTML files that are explicitly intended to be shared with non-developers via email. This creates a data exposure risk if the generated file contains sensitive business logic, API structures, or internal state data that should not leave the secure project environment.
[COMMAND_EXECUTION]: The UI prototype workflow (UI.md) relies on launching code via project task runners like pnpm, python, or bun. This creates an execution surface where the agent is prompted to run local scripts, which requires careful oversight to ensure only intended code is executed.
[DYNAMIC_EXECUTION]: The skill's primary function is to generate and potentially integrate executable HTML and UI components into a codebase. While it includes instructions to rewrite code to production standards, the direct integration of logic modules from prototypes into production files is a dynamic execution risk if not subjected to thorough human review.

prototype