Supplementary Material: Agent Architecture and Phase-Specific Prompts

This resource documents the complete prompt set of a human-in-the-loop LangGraph agent that co-develops supervised regression pipelines along CRISP-DM. Each LLM-calling phase pairs a stable system prompt - fixing role, output contract, sandbox conventions, judgment calls, failure modes, and scope boundaries in labelled sections - with a user-prompt template that injects run-time context through placeholder variables: the current DataFrame columns with dtype, cardinality and missing-value share, a compact cross-phase project memory, and the concrete task specification. Roles are assigned consistently: senior data scientist for interpretive phases, senior data engineer for code-generating phases. All outputs are constrained by Pydantic schemas, and shared building blocks (e.g. the follow-up-question policy) are injected from a single source to avoid divergence. Three data-preparation phases share one code-generation prompt; feature engineering uses a three-call design separating code generation, post-hoc documentation, and a failure critic whose diagnosis is fed back into a bounded self-correction loop of at most five attempts. Deterministic phases - splitting, encoding, model training, evaluation - contain no prompts at all. Prompts are versioned per phase and reproduced verbatim in German, the language in which they are sent, with English translations for readability.