The Agent Skills Directory

[SAFE]: The skill provides logic via a Python zipapp (scripts/common.pyz) which bundles several utility modules including cli.py, findings.py, and handoff.py. This is a standard distribution format for Python command-line tools and no suspicious or malicious behavior was found in the surrounding instructions.
[SAFE]: The prompt templates (e.g., curd-prompt.md, spawn-primitive-reference.md) define a structured multi-agent workflow with clear role boundaries and permission enforcement. The system explicitly requires isolated worktrees and fresh contexts for sub-agent tasks, which aligns with security best practices for autonomous agent architectures.
[SAFE]: The skill incorporates a spec decomposition workflow that processes external inputs. An analysis of the indirect prompt injection surface was conducted:
Ingestion points: User-provided specifications are ingested via the references/decomposer-prompt.md template.
Boundary markers: The templates utilize Markdown delimiters and explicit scope constraints (e.g., "Touch ONLY the named file", "Stay inside this behaviour") to separate data from instructions.
Capability inventory: The workflow utilizes local scripts (common.pyz) and project tools (/plate) for structured metadata management and quality gate execution.
Sanitization: Output artifacts are validated against defined JSON schemas (references/manifest-schema.json) to prevent the propagation of malformed or malicious data between workflow phases.

ultracook