How Minsheng Bank Applies Spec‑Driven Development with CodeAgent for Private‑Domain R&D

Recent advances in large‑model technology have deepened AI’s role in software development, moving from simple code‑completion tools to assistants that can understand requirements, generate designs, write code, and aid testing. The bank initially used CodeAgent for rapid prototyping, but strict banking regulations, complex systems, and the need for consistent engineering standards revealed that conversational AI alone could not guarantee accurate, compliant code.

In 2025 Minsheng Bank’s technology team launched an exploration of Spec‑Driven Development (SDD). Leveraging a privately deployed Cloud IDE and the Minsheng Code CLI, they adopted a workflow where enterprise‑level, domain‑level, and project‑level specifications drive AI‑generated code that conforms to private‑domain standards and functional needs.

Background and challenges identified include AI’s limited global code understanding, difficulty generating code that meets functional requirements from human‑focused PRDs, poor adherence to private‑domain norms, lack of stable quality‑verification mechanisms, and difficulty reusing development assets.

Understanding SDD references the September 2025 GitHub proposal that treats a Spec as a contractual code‑behavior guide. ThoughtWorks engineer Birgitta’s three‑level SDD strategy—Spec‑first, Spec‑anchored, and Spec‑as‑source—is highlighted. The bank defines specifications as clear, verifiable, evolvable, AI‑readable technical descriptions, categorized into enterprise‑level (technical policies, framework rules), domain‑level (business models, engineering requirements, middleware usage, testing guidelines), and project‑level (functional specs, interface contracts, data structures, acceptance criteria).

The SDD workflow is abstracted into five stages: Spec → Plan → Task → Implement → Verify . Each stage is described: specifications clarify requirements; planning defines implementation paths; tasks break down work; implementation covers code generation and system building; verification ensures outputs meet acceptance criteria.

Practical experience revealed early issues such as mismatches between existing agile/waterfall artifacts and ready specifications, added burden on developers to interact with AI, and inconsistent spec quality across teams. After iterative adjustments, the team standardized spec templates, layered specifications, and reinforced constraints and acceptance conditions. The practice yielded measurable improvements: generated code reached a quality baseline with unit‑test coverage near 70%, compliance with private‑domain standards exceeded 70%, and developers could shift from coding to higher‑value tasks like code review.

Key SDD patterns include injecting organization‑level information (enterprise specs) as AI context, reusing domain‑level knowledge assets, focusing on project‑level specs for concrete implementations, strengthening constraints and acceptance criteria (e.g., prohibiting certain methods and generating accompanying unit tests), and extending the workflow with AI Skills and MCP for tasks such as architecture validation and test generation.

Limitations acknowledge that SDD is not suitable for all project types (e.g., exploratory prototypes, UI design, algorithm research, financial‑logic intensive, performance‑tuning, or minor changes). Over‑specification can create a maintenance burden as specs diverge from code, and long‑term spec asset freshness requires versioning and contextual management tools.

Future actions involve aligning SDD with Harness Engineering principles, introducing multi‑agent coordination for boundary checks and test augmentation, and redefining AI‑coding metrics beyond traditional code‑adoption rates to focus on quality and efficiency outcomes.

In conclusion, as large‑model capabilities evolve, AI‑enabled development must operate within disciplined, spec‑driven processes to achieve stable, efficient, and compliant software delivery in banking. Minsheng Bank’s ongoing SDD practice demonstrates a pathway for integrating AI into private‑domain R&D while continuously refining specifications, tooling, and measurement approaches.