AI Case Analysis: Submission Guidelines
AI research is genuinely valued here — including when patients use it to synthesize their own cases. The problem is unformatted AI output: often 20+ pages, most of which is extraneous for a clinician with advanced training. To be useful at the visit level, it needs to be reformatted first. The prompt below does that.
Before you start: Read the blog article on AI limits at https://docsandford.com/archives/articles/ai-confabulation-in-healthcare. Understanding where AI goes wrong will help you interpret its output — and help us use it more effectively together at your visit.
One heads-up about the last step: The prompt ends with a second-AI cross-check. A second AI will always find something — that’s expected. Fix genuine clinical errors or major overconfidence flags; note minor quibbles and stop iterating. One or two rounds is enough; bring anything unresolved to the appointment.
Prompt to provide the AI:
“Please recreate this entire analysis for my integrative medicine physician, who has advanced medical training. Format it exactly as follows:
- TL;DR — 2 to 4 sentences maximum. Synthesize the entire case into tight clinical context. Lead with the patient’s primary complaint and goal for this visit — not what is most interesting, but what is most relevant. Include the AI model name, version, and self-assessed intelligence or capability level (you may need to ask yourself this directly).
- Tight clinical bullet points expanding on the TL;DR. Assume the reader does not need background pathophysiology, definition of lab values, or general medical education. No redundancy with other sections. This entire section must fit within 6 inches of vertical space on a standard 8.5×11 page — enforce this strictly.
- Red flags and urgent clinical considerations first — surface any drug interactions, contraindications, urgency tiers, or safety concerns at the very top of the bullet section, before other content.
- Assessment — a mechanism-based clinical synthesis, not a list of possibilities. Rank impressions by confidence and state your reasoning. Explicitly distinguish between what the patient reported, what the labs or records show, and what you are inferring. Label inferences as inferences.
- Derived questions, suggested testing, and potential treatment ideas — be specific. Flag any treatment ideas that are speculative, evidence-limited, or extrapolated beyond the literature. For any factual medical claim, cite a source or explicitly state that the claim is unsupported and based on reasoning only.
- Knowledge currency notice — flag any area where your training data cutoff may affect the relevance or accuracy of your guidance.
- Supporting detail — any further information that didn’t fit above. Keep this section as short as possible. No redundancy with prior sections.
- Prompt log — list every prompt provided to you during this conversation, verbatim and in sequence, so the physician can assess the quality and direction of the inquiry. Do not paraphrase.
- Self-audit — now review everything you just produced. Assume a skeptical clinician will challenge every claim. Identify any assertions that are unsupported, any inferences presented as facts, any internal inconsistencies between sections, and any areas where your confidence is lower than your language implies. List these explicitly.”
Recent Comments