Artificial intelligence has rapidly become part of many physicians' daily workflows, helping draft documentation, summarize medical literature, and answer clinical questions. Yet for endocrinologists, one persistent concern remains: How to trust recommendations generated by a general-purpose large language model (LLM) when the source of reasoning is often unclear?  

For Om J. Lakhani, MD, DNB, consultant endocrinologist at AB Plus Hospital in Ahmedabad, India, the solution is not a more powerful LLM but better structured clinical guidance. That idea led to MD2SKILL, an open-source project that converts evidence-based clinical guidelines into structured instruction files (called Skill.md) that artificial intelligence (AI) assistants load when a patient's presentation matches the relevant clinical scenario. 

“Rather than relying primarily on the model's internal knowledge or internet search, the system directs the AI through a predefined clinical workflow while citing the original source,” says Dr. Lakhani.  

Technology mindset meets clinical practice  

Dr. Lakhani's interest in AI began long before ChatGPT. "I taught myself coding in an era when it wasn't very popular," he said.

As LLMs matured, he began experimenting with early AI systems and developed DoctorGPT to assist with manuscript drafting and patient communications. The idea for MD2SKILL, however, emerged from his everyday clinical practice.  

Dr. Lakhani estimated that he sees 30 to 40 patients each day. Many encounters require rapid decisions about increasingly complex clinical questions. For example, does a child meet criteria for genetic testing for monogenic obesity? Does this thyroid nodule require fine-needle aspiration?  

According to Dr. Lakhani, most of these answers exist—but are often buried within comprehensive clinical practice guidelines spanning dozens or even hundreds of pages. "The guidelines are excellent, but they are also entirely unread at the moment of the consultation, when you need them the most," he said.

Rather than having AI summarize these lengthy guidelines, Dr. Lakhani asked a different question: "What is the smallest practical piece of a guideline that could help a physician make a bedside decision in less than a minute?"

The answer became MD2SKILL. "Instead of an AI model answering clinical questions pulled from its own knowledge or the Internet, the MD2SKILL system gives you answers directly tied to guidelines from the major specialty societies, the ones a practicing endocrinologist already treats as authoritative," said Dr. Lakhani. “It is the AI equivalent of a tightly-written clinical pearl—the kind a senior consultant would hand a junior on a busy ward.”  

MD2SKILL at work  

To illustrate how MD2SKILL functions in clinical practice, Dr. Lakhani shared a recent case from his own experience.  

A 19-year-old male patient had a body mass index of 46 kg/m² with progressive weight gain, night blindness, polydactyly, a history of syndactyly correction and orchidopexy, and consanguineous parents. 

Dr. Lakhani suspected a possible syndromic or monogenic cause for the obesity. He described the case to Claude—the AI platform currently used by MD2SKILL—and it pulled two Skill.md files from its database—"monogenic-obesity-diagnosis" and "nhs-genomic-test-finder"—and presented a clinical reasoning chain:  

• Diagnosis: Bardet-Biedl syndrome (BBS), near-textbook constellation.  

• Cardinal features: rod-cone dystrophy, polydactyly, central obesity, hypogonadism. Renal involvement flagged for screening.  

• Differential: polydactyly rules out Alström syndrome, which otherwise overlaps, and favors BBS.  

• Workup: NHS gene panel (specific test code returned by the second skill), clinical exome as an acceptable alternative.  

• Treatment implication: setmelanotide is licensed for BBS-related obesity once a confirming gene variant is in hand. Without the gene, no setmelanotide.  

In this case, the information helped guide Dr. Lakhani's next clinical steps, which included genetic counseling for the family and confirmatory genetic testing. "Only after a confirmed variant did we discuss prescribing setmelanotide," he says. "The Skill.md file doesn't make that decision for me. It makes sure I'm asking the right question and ordering the right test the first time, instead of a round of trial-and-error investigations." 

(For a step-by-step video tutorial on using MD2SKILL, click here.) 

Consistency in design  

Skill.md files are hosted on md2skill.org and in the project’s public MIT-licensed GitHub repository. As of June 2026, MD2SKILL contains approximately 305 Skill.md files across 11 specialties, including 271 focused on endocrinology. Topics include pheochromocytoma, lipid disorders, primary aldosteronism, inpatient diabetes, metabolic bone disease, glucagon-like peptide-1 receptor agonists, diabetes in pregnancy, thyroid cancer, and preventive endocrinology. The project is expanding into cardiology, neurology, urology, and pediatric endocrinology. Clinicians also can now fork existing Skill.md files and contribute new ones to the GitHub repository. 

Every Skill.md file follows a standardized template designed for consistent, transparent, and rapid clinical use. Each file includes: 

Trigger descriptions. The description identifies the clinical scenario in language physicians are likely to use. “If the description does not contain the actual clinical phrases a doctor would use—for example, ‘diabetic foot infection,’ ‘high-dose tirzepatide,’ or ‘ICU adrenal insufficiency’—Claude will never load it,” said Dr. Lakhani. 

Structured decision pathways. Rather than presenting a list of facts, the file organizes guidance in the sequence a physician would follow when evaluating a case. “The pathway is the order in which a clinician has to think through a case—not just a list of facts, but a sequence of decisions where each step narrows things down for the next one,” said Dr. Lakhani. 

Clinically actionable recommendations. Each file includes only information that affects evaluation or management and excludes unnecessary background material. 

Source citations. Recommendations link directly to the original guideline, allowing physicians to verify the source and independently assess the guidance. “If a clinician disagrees with the output, they can open the file, read the source, and decide for themselves,” said Dr. Lakhani. 

Rapid consultation workflow. Each skill is intended to be used in less than 60 seconds during a consultation. Files exceeding 100 lines of core content are divided into supporting components: extensive drug tables are moved to a REFERENCE.md file, worked cases to an EXAMPLES.md file, and deterministic calculators to a Python script. 

Clear instructions. Skills are written so that clinicians can understand and apply them after a single reading. If not, the file is rewritten. 

Building each Skill.md file  

Every Skill.md file undergoes a structured review process before being uploaded. First, Dr. Lakhani and his colleague, psychiatrist Archit Mehta, MD, who serves as the project's technical co-developer, convert clinical practice guidelines into a draft Skill.md file. The team then checks every recommendation, line by line, against the source. "For any topic outside our own direct expertise, an outside specialist reviews it," said Dr. Lakhani.  

Different professional societies occasionally issue recommendations that vary in emphasis or timing. Evidence can also be limited, conflicting, or absent. Rather than allowing AI to reconcile those differences independently, Dr. Lakhani developed a separate workflow to identify areas of disagreement, enabling outside clinicians to review and reach a consensus. When recommendations differ between older and newer guidelines, the team generally adopts the newer recommendations while preserving citations to the original guideline documents.  

The development process has also yielded unexpected insights. “On several occasions, the AI identified relationships or recommendations within guideline tables and figures that the investigators themselves had initially overlooked,” according to Dr. Lakhani.   

Challenges and future research

Dr. Lakhani recognizes that the project is not without its challenges. His greatest concern is the risk of hallucinations—false, misleading, or fabricated information generated by an AI model.  

"Vetting the source guideline doesn’t remove the risk, because the model still has to compose the final answer—it’s blending the Skill.md file's grounded content with its own generative reasoning in real time," he noted. "A perfectly accurate file can still be misapplied outside the clinical scenario it was written for, or the model can quietly blend in something from its own background training that was never in the vetted source."  

Whether the Skill.md framework reduces hallucinations and improves guideline fidelity remains an open question that the team hopes to answer through a prospective blinded validation study.  The study will compare Skill.md-assisted responses with those generated by a generic LLM using the same clinical cases. Endocrinologists, blinded to how the responses were generated, will evaluate their quality. 

"If we can show that this approach performs better than a generic LLM, it would support the idea that a systemic approach can produce better answers, especially in health care where precision is very important and necessary," said Dr. Lakhani.  

By emphasizing structured guidance, transparent reasoning, and direct links to evidence-based recommendations, MD2SKILL represents one emerging approach to making AI a more reliable clinical decision-support tool—an approach that now awaits prospective validation.

(Dr. Lakhani disclosed that has no current commercial interest in MD2SKILL, adding that the program is offered as a free, open resource for the clinical community.)