llm-instruction-separation

What problem does it solve?

Prevents hard-coded model identifiers and long LLM instruction templates from being embedded directly in Wolfram Language (.wl) sources, improving maintainability and routing correctness.

Core Features & Use Cases

• Model capability separation: Ensures model “branches” (e.g., gpt-5, claude-opus-4.7) live only in the $ClaudeModelCapabilities table keys, not in code conditionals.
• Prompt/template extraction: Moves LLM instruction text (e.g., $petriNetGuideExtras) into skill Markdown so .wl loads it via a dedicated reader.
• Rules vs. skills discipline: Keeps cross-package operational conventions (quiet/check behavior, working directory guidance) in rules, while package-specific techniques stay in dedicated skills.
• Migration guidance: Provides before/after patterns for cutting string constants out of .wl into skills, including safe directive-root resolution.

Quick Start

Ask your AI to scan your .wl package for hard-coded model names and long LLM instruction strings, then refactor the long instructions into a new skill named llm-instruction-separation and update the .wl code to load the skill body via iReadSkillBody.