The Agent Skills Directory

[SAFE]: The skill contains no executable code, network operations, or persistence mechanisms. Its behavior is limited to text-based analysis of biomedical research findings.
[PROMPT_INJECTION]: The instructions include defensive guardrails such as a mandatory out-of-scope redirect (Step 1) and 15 hard rules that prevent the model from generating clinical recommendations or fabricating scientific evidence.
[INDIRECT_PROMPT_INJECTION]: The skill provides a surface for ingesting user-provided biological discovery data, but it lacks dangerous capabilities in the execution environment. The evidence chain is documented as follows:
Ingestion points: User-provided research findings and discovery units in SKILL.md.
Boundary markers: Explicit out-of-scope templates and 'Hard Rules' that distinguish research analysis from clinical advice.
Capability inventory: No subprocess calls, network access, or file-write capabilities are requested or used.
Sanitization: Logical validation and narrowing of inputs are required in Step 1 before analysis proceeds.
[METADATA_POISONING]: The skill metadata and author information are consistent with its stated purpose. The included evaluation report (eval_report_...json) provides development context but does not contain malicious instructions or deceptive labels that compromise the auditing process.

basic-discovery-translational-opportunity-finder