Task Boundary Auditor
Paste the following prompt into your AI chat to install this skill:
Please install @user_15292d5a/yjkj-task-boundary-auditor according to https://skillhub.cn/install/skillhub.md.
About this skill
Problem
LLM boundary violations often do not surface as errors. Instead, the model returns a well-formatted, coherent answer that is still fundamentally wrong. This is the “sophisticated hallucination” risk: counterfactual reasoning, zero-omission auditing, precise verification, real-time closed-loop control, or physical intuition can all produce polished outputs without the required causal model, formal tools, or human accountability. The skill performs a pre-execution audit to prevent high-risk subtasks from being assigned to the LLM as the primary executor.
How It Works
The skill classifies tasks into Safe Zone, Restricted Zone, and Prohibited Zone, then audits them using detection signatures:
- Prohibited Zone: counterfactual reasoning, precise verification, real-time closed-loop, physical intuition, conceptual extrapolation, ethical adjudication, zero-omission audit, and extreme prediction.
- Safe Zone: text generation, translation and conversion, information extraction, summarization, and formatting.
- Restricted Zone: short-chain causality, numerical estimation, draft proposals, and known pattern matching.
For mixed tasks, it decomposes the request into subtasks, classifies each one separately, and outputs a violation block, restricted routing, or safe clearance result.
Boundaries
This skill is useful for auditing high-risk requests, decomposing complex tasks, and labeling what the LLM can and cannot own. It does not replace professional tools, rule engines, statistical models, or human judgment. In legal, medical, financial, or other professional domains, human review should still be recommended even when the task is technically feasible. Audit conclusions must not be downgraded because of user urgency, vague phrasing, or pressure for a final answer.
Use Cases
- Audit a request to find all hidden clauses in a 200-page contract and block zero-omission guarantees.
- Decompose sales analysis into factor extraction, statistical-model judgment, and historical description.
- Clear a safe task to translate a contract and extract payment clauses, with spot-check notes.
- Flag counterfactual “what if it never happened” reasoning as prohibited and route it to expert analysis.
Best For
- AI Agent safety engineers auditing high-risk prompts and routing human review.
- Operations staff handling contracts or compliance requests needing a clear LLM/human split.
- Platform engineers building support or knowledge-base Agents with task routing rules.
- Evaluation engineers reviewing model outputs with a consistent boundary classification table.
Related Skills
Turn vague requests into structured prompts with role, task, context, and constraints, outputting only usable prompts without executing the task.
A skill for game NPC and enemy AI that separates decision, steering, and pathfinding, covering FSMs, behavior trees, A*, path caching, and pitfalls.
Provides multi-turn emotional companion chat through a third-party Agent API and supports continuing sessions via session_id.
An evaluation-driven workflow for diagnosing, optimizing, validating, and logging Agent Skills based on design patterns and anti-patterns.