- The paper introduces vibe coding, where clinicians leverage LLMs as programming partners to rapidly prototype tailored digital health applications.
- It outlines three distinct paradigms—conversational LLMs, integrated platforms, and agentic coding—supported by a comprehensive risk stratification framework.
- The study demonstrates practical case studies and governance measures that balance rapid prototyping with the need for quality and regulatory oversight.
Vibe Coding for Clinicians: Software Co-Development via LLMs in Digital Health
Introduction
"Vibe coding for clinicians: democratising bespoke software development for digital health innovation" (2604.22604) formulates and interrogates the concept and practical impact of "vibe coding"—the LLM-assisted co-development of software using natural language prompts—within the medical context. The authors, who are clinical early adopters spanning a range of technical proficiency, outline a rigorous yet pragmatic pathway for clinicians to leverage LLMs as direct programming partners, lowering barriers to problem-focused tool prototyping without formal software engineering backgrounds.
Motivation and Context
The core driver addressed is the "long tail" of health IT needs: operational bottlenecks, research logistics, and educational requirements considered too context-specific or low in perceived impact to trigger traditional commercial interest. Current EHRs and clinical informatics platforms are optimized for scale, not nuanced usability, resulting in widespread inefficiency and occupational burnout. Access to programming talent or development funding for local or highly-personalized solutions is rare. The authors assert that the generative competence of LLMs, channelled via natural language, enables direct, domain-expert-led prototyping, mitigating chronic developer-clinician misalignment.
Vibe Coding: Workflow, Taxonomy, and Technical Principles
Three paradigms of vibe coding are delineated:
- Conversational LLMs as Programming Partners: LLMs such as GPT-5, Claude, or Gemini are prompted iteratively in natural language, with generated code manually integrated and tested by the clinician in a local IDE. This offers maximal transparency but requires technical mediations (context management, error resolution) by the user.
- Integrated Platforms: Cloud-based, AI-native IDEs (e.g., Lovable, Replit) streamline the path from prompt to executable MVP, abstracting deployment, live testing, stack configuration, and even front-end design. These platforms reduce cognitive and technical overhead but introduce vendor lock-in and often limited portability.
- Agentic Coding: Advanced systems (e.g., Claude Code, Devin) frame the LLM as an autonomous, goal-directed entity capable of multiphase build-refine cycles, with clinicians specifying high-level requirements. This has greatest utility for developers with moderate to advanced engineering proficiency.
The recommended workflow for non-engineer clinicians adheres to a structured iterative loop, consistent with contemporary agile methods: explicit vision and scope definition, translating requirements to context-rich prompts, iterative code testing and debugging via the LLM, and strict discipline around documentation and version control. The manuscript strongly cautions that while LLMs enable code generation, quality, maintainability, and technical debt remain substantial risks in the absence of foundational software development literacy.
Critical Risk Stratification and Governance
A pivotal contribution is the explicit risk stratification framework for vibe coded clinical tools, scaled across three tiers with commensurate validation and governance requirements:
- Tier 1 (Personal): Tools with no clinical influence or patient data, governed solely by the developer's self-testing and prompt/code documentation.
- Tier 2 (Administrative): Indirectly impacting clinical workflows, requiring peer review, maintenance assignment, and notification of local IT/governance, especially if networked.
- Tier 3 (Patient-Facing/Clinical Support): Tools that process PHI or contribute to clinical decision-making, mandating formal validation, professional software development input, institional and potentially regulatory approval (e.g., as SaMD).
This taxonomy provides a nuanced operational guideline, addressing the spectrum from informal "shadow IT" to clinically consequential systems, with explicit escalation of due diligence as usage matures or broadens.
Case Studies and Empirical Examples
The manuscript details archetypal implementations:
- A custom calculation tool for ophthalmology injection intervals, built and deployed personally, exemplifying high-frequency, low-stakes workflow facilitation achievable by non-experts.
- The "Cataract Calendar" postoperative eye drop adherence tool, progressing from a hackathon prototype to institutional pilot testing. This demonstrates the framework's support for migration across risk tiers with scalability in governance and validation.
Such cases substantiate the practical capability of clinician-driven vibe coding to eliminate persistent workflow inefficiencies and address critical adherence and safety gaps.
Theoretical and Practical Implications
The authors underscore that vibe coding does not diminish the necessity for professional engineering as solutions scale or are deployed at scale. Rather, the key impact lies in enabling clinicians to serve as hybrid user-innovators: rapidly iterating needs-driven prototypes, uncovering hidden dependencies, clarifying requirements, and thus optimizing subsequent handover to technical teams. There is significant emphasis on the augmentation of clinical agency and the improved communication between end users and engineers by virtue of working MVPs.
The discussion extends to intellectual property management, regulatory compliance, and the risks of license-infringing code fragments—a non-trivial consideration given empirical findings that 0.88–2.01% of LLM output may mirror licensed open source [see Xu et al., (Xu et al., 2024)]. The use of prompt/version documentation, software bills of materials, and pre-deployment compliance audits is recommended.
Limitations and Forward Directions
Vibe coding presents inherent risks of unchecked technical debt, maintenance fragility, and security vulnerabilities—particularly acute for users with limited software engineering awareness. The risk of overreliance on LLM-generated code and hallucinations remains, particularly where there is domain-expertise asymmetry between the user and the underlying model. The authors urge the promotion of software literacy and disciplined prompt engineering as necessary conditions for safe use.
The implications for the digital health innovation lifecycle are substantial. The prospect of dynamic, end-user-driven tool pipelines (with appropriate regulatory/engineering checkpoints) points toward a flattened, more responsive health IT innovation ecosystem. The paradigm may extend to other knowledge-intensive sectors where domain experts lack programming skills but now gain direct access to LLM-augmented software creation.
Conclusion
The manuscript provides a comprehensive, technically sound framework for the integration of LLM-driven co-development ("vibe coding") into clinician-led digital health innovation. The approach fosters domain-expert empowerment, accelerates translational prototyping, and compresses the communication loop between clinicians and engineers, provided that robust governance frameworks are employed and software development fundamentals respected. The practical and theoretical significance lies less in the automation of engineering and more in the democratization of ideation and prototyping at the clinical frontier.