Developer Rebuilds Text-to-Diagram AI Tool as Multi-Turn Agent to Fix Key Flaws

A developer overhauled their AI-powered text-to-diagram tool after its single-LLM-call architecture repeatedly produced syntax errors and silent semantic drift in production. The updated version, called v2, operates as a multi-turn agent with two distinct loops: an outer loop that clarifies user intent and an inner loop that generates and validates Mermaid diagram code. The system is modeled as a law firm pipeline, with six specialized roles — Clerk, Auditor, Planner, Lawyer, Reviewer, and Front Desk — each handling a single responsibility to improve debuggability and reliability. A central orchestrator called Front Desk maintains conversation state and sequences the roles, while the Planner serves as the sole decision point for whether the agent should ask the user a question or proceed to generate a diagram. Splitting responsibilities across isolated roles replaced an unwieldy single mega-prompt and made each component independently testable.