The Agent Skills Directory

[SAFE]: The skill is a comprehensive framework for repository organization and agent alignment. It implements a multi-phase workflow (Assess, Plan, Knowledge, Architecture, Enforcement, Quality, Process, Verify) to build a 'harness' for AI agents.
[METADATA_POISONING]: The metadata correctly identifies the organization as 'OpenAI / Harness Engineering' and provides relevant, trusted references (openai.com, cookbook.openai.com). The stated purpose and organization match the technical content provided.
[INDIRECT_PROMPT_INJECTION]: The skill processes repository content to generate reports and configuration files. While this is an ingestion surface for untrusted data, the skill includes explicit instructions for human steering and review at each phase boundary ('present results and waiting for user confirmation'). The capabilities are limited to generating documentation, configuration files, and lint/test stubs, which are central to the skill's stated purpose.
[DYNAMIC_EXECUTION]: Phase 5 involves generating artifacts like lint configurations, CI workflows, and structural tests. These are generated based on the architectural assessment and golden principles identified during analysis. The instructions emphasize that these artifacts should serve as enforcement tools for the harness, and no unsafe code generation or interpolation patterns were found.
[EXTERNAL_DOWNLOADS]: No suspicious external downloads or remote code execution patterns were detected. References point to well-known and trusted technology domains (OpenAI, GitHub community blogs).

harness-engineering