Dianoia AI: An Innovative Framework for Distilling Domain Expert Knowledge into AI-Injectable Skill Files

Dianoia AI is an open-source project proposed by the bigmoon-dev team, aiming to extract and encapsulate domain experts' reasoning patterns into standardized skill files (profile.yaml and SKILL.md), enabling knowledge transfer and injection into any AI model. Addressing the pain points of domain adaptation for general-purpose LLMs, this framework provides a new knowledge transfer method that does not rely on fine-tuning or RAG; instead, it captures the procedural knowledge of how experts "think", featuring model agnosticism and flexible reusability.

Current large language models (LLMs) have strong general capabilities but underperform in specific professional domains. Traditional domain adaptation methods like fine-tuning (high cost, requires large amounts of labeled data) and Retrieval-Augmented Generation (RAG, limited by retrieval quality and context window) have limitations. Dianoia AI takes a different approach, exploring a new paradigm of extracting experts' thinking patterns and reasoning paths into structured skill files.

"Dianoia" originates from ancient Greek philosophy, meaning "thinking" or "reasoning". The project focuses on capturing experts' mental models and reasoning chains for problem-solving. Unlike traditional methods, it distills procedural knowledge of "knowing how to think" rather than factual content. It adopts a dual-file architecture: profile.yaml defines expert role characteristics, knowledge boundaries, etc. ("expert persona") in a structured format; SKILL.md describes skill operation processes, key considerations, etc. ("operation manual"). Both files can evolve and be reused independently.

The knowledge extraction process includes: analyzing domain corpora (technical documents, papers, etc.) to identify reasoning patterns → abstracting and generalizing into structured representations → manual review and refinement. The generated skill files use a standardized format and can be injected into any AI model that supports system prompts (GPT-4, Claude, Llama, etc.), avoiding vendor lock-in. Users can flexibly choose models while maintaining consistent professional capabilities.

Application scenarios include law (inheritance of case analysis thinking), healthcare (diagnostic reasoning assistance), software development (architectural decision guidance), etc. It is suitable for scenarios that require inheriting tacit knowledge and shortening the training cycle for new employees. Comparative advantages: Compared to RAG, it captures deep reasoning knowledge rather than factual retrieval; compared to fine-tuning, it does not require expensive resources or labeled data, and knowledge updates are more flexible (just edit the files).

Challenges faced: Difficulty in explicitizing expert reasoning, skill file quality depending on source corpora, and the need to establish an evaluation system to verify model outputs. As an open-source project, it is expected to form a skill file sharing ecosystem in the future, where experts from various fields contribute skills, and developers can quickly install professional skills for AI, promoting the transformation of AI from general-purpose to deep professional capabilities.