The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external natural language specifications (nlspec) and existing codebase content to generate implementation plans and work items.
Ingestion points: The skill instructions specify 'nlspec' and 'current code' as the primary inputs for the slicing process.
Boundary markers: The skill establishes a 'handoff contract' and relies on specific artifacts like the 'slice spec' to bound the scope of work. It does not explicitly mention technical delimiters but uses the 'nlspec' as a governing document.
Capability inventory: The agent using this skill performs planning, technical specification writing, and generates code diffs and file-tree modifications.
Sanitization: The methodology incorporates strong mitigation strategies, specifically requiring Human-In-The-Loop (HITL) review for 'security-class cells' (handling credentials, authentication, or authorization) and any seam-touching architectural changes. It also mandates that the model performing the work cannot adjudicate its own compliance, requiring independent review.
[SAFE]: The skill explicitly addresses security risks by identifying 'tripwires' and requiring human oversight for sensitive operations such as credential handling and data migrations. It acknowledges the possibility of rule-gaming and implements multi-model review as a countermeasure.

slice-design