Can codified expertise let non-experts match specialist output?

Critical domain knowledge usually resides with a few experts, creating organizational bottlenecks: when experts are unavailable, work halts or proceeds with suboptimal outcomes, and experts must trade their primary work against mentorship. The paper studies this through simulation data visualization, where non-experts default to familiar chart types because choosing appropriate techniques for complex data is hard, and even attempted sophisticated visualizations need expert interpretation.

The contribution is a software-engineering framework for capturing and embedding human domain knowledge into an LLM agent — not a single prompt but a composed system: a request classifier, a RAG system for domain-specific code generation, codified expert rules, and visualization design principles, unified in an agent exhibiting autonomous, reactive, proactive, and social behavior. Across five scenarios with twelve evaluators it delivered a 206% output-quality improvement, reaching expert-level ratings in all cases (versus the baseline's poor performance) with superior, lower-variance code quality.

The keeper claim is organizational: codifying tacit expertise into an agent's scaffolding dissolves the expert bottleneck — non-experts produce expert-level outputs without consuming expert time. The mechanism is that expertise lives in the rules and design principles deliberately externalized into the harness, not in the base model's general capability. This is the single-domain, knowledge-codification cousin of Does structured artifact sharing outperform conversational coordination? — both show that codifying human procedure into structured agent scaffolding beats leaving it implicit.