What it does
This skill acts as a specialized static analysis tool for LLM integrations, specifically designed for the "Fable 5" era of AI. It scans your Python or TypeScript source code to identify call sites that assume a model will always return a successful, valid response. It flags instances where code lacks 'refusal' handling or fails to account for silent fallbacks to older model versions (like Opus 4.8) when a high-risk request is blocked.
Why use this skill
Standard linters don't understand LLM lifecycle risks. As safety guardrails become more prevalent, your code is increasingly likely to receive a "refusal" or a response from a less-capable fallback model. If your code parses these responses blindly, it will crash or produce degraded results. This skill identifies these "blind spots" so you can implement graceful handling before they hit production.
Supported tools
- Languages: Python, JavaScript, TypeScript
- Frameworks: Common LLM client patterns (OpenAI, Anthropic, LangChain)
- Workflow: CLI-based scanning with remediation snippets provided
The output provides a detailed report including rule IDs, severity levels, and specific lines of evidence, making it easy to integrate into your CI/CD audit process.