The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied marketing briefs and evidence to generate strategic recommendations. While this presents an ingestion surface for potentially untrusted data, the skill implements a comprehensive "Evidence discipline" and "Operating contract" that mitigates injection risks.
Ingestion points: User-provided marketing data and objectives as defined in SKILL.md and simulated in evals/behavior.yml.
Boundary markers: The references/output-contract.md enforces a structured response format that separates evidence from inference, and the SKILL.md instructions explicitly mandate reading the skill-spec.json invariants before processing.
Capability inventory: The skill uses local routing scripts (dynamic_router.py, neural_router.py) and optional web research tools.
Sanitization: Strong evidence policy requires tagging claims as fact, inference, or assumption, and explicit invariants in references/skill-spec.json forbid the fabrication of customer language or market facts.
[COMMAND_EXECUTION]: The skill references local orchestration scripts (e.g., ../../scripts/dynamic_router.py, ../../scripts/neural_router.py) to manage complex workflow dependencies and routing. These are internal vendor resources intended for agent coordination within the framework and do not involve executing untrusted or remote code.
[SAFE]: The skill's metadata and instructions are consistent with its stated purpose. It includes extensive evaluation files (evals/*.yml) to ensure the agent adheres to safety and behavioral constraints, such as refusing to fabricate certainty when evidence is missing.

category-strategy