The Agent Skills Directory

[SAFE]: The skill defines an iterative refinement process for AI-generated code and prompts. No malicious code execution, network exfiltration, or unauthorized file access patterns were identified.
[PROMPT_INJECTION]: The instructions establish a rigorous framework for evolving agent behavior through reviewed failures. It does not contain patterns attempting to bypass safety filters, override system prompts, or extract sensitive instructions.
[DATA_EXPOSURE]: While the skill includes templates for artifact registers and repository paths, it does not access sensitive local files (such as SSH keys or credentials) or perform network requests to external domains.
[INDIRECT_PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by processing external failure data (counterexamples) to modify executable code.
Ingestion points: External simulation context and counterexample traces are ingested via the forms in references/working-forms.md.
Boundary markers: The skill mandates strict acceptance rules (accepted(case, P, S)) and review protocols to separate implementation defects from authorized policy changes.
Capability inventory: The agent is tasked with compiling executable "projections" (code, prompts, or configuration) in SKILL.md.
Sanitization: The framework relies on a human or model authority to review simulated output against a fixed sketch and requires deterministic regression gates to prevent unauthorized behavior changes.

cess