Five Questions Before the First Word
A 140-line skill that interviews me before it writes, and why the pattern belongs in clinical practice.
Five Questions Before the First Word
A 140-line skill that interviews me before it writes, and why the pattern belongs in clinical practice.
The short version: I installed a 140-line skill that makes Claude interview me before it writes anything substantive. It asks the three to seven questions whose answers would most change the output, proposes its best guess for each, and refuses to draft until I approve the brief. The discipline is the oldest one in medicine: interrogate the chief complaint before treating it. The same pattern improves patient handouts, workflow memos, noon-conference talks, and murky-case thinking. It ports to any model in 43 words.
It is 8:40 on a Tuesday evening and I am at my desk in Plano, typing a two-sentence request into Claude. I want a build spec for a small tool that drafts disability paperwork from chart data, the kind of project that eats a weekend if I start it wrong. I expect a draft. I have made requests like this a hundred times, and a hundred times the model has obliged with something competent, generic, and 70 percent right.
This time the model does not write a word. I have a skill installed called the interviewer, and the skill intercepts the request. A skill, in Claude’s vocabulary, is a file of standing instructions the model loads when a task matches it; this one tells the model that substantive creation work begins with questions, not output. What comes back instead of a spec is a working outline with the gaps marked, followed by five questions. Each question arrives with the model’s best-guess answer attached, so I can confirm or redirect rather than compose from a blank page.
Question three stops me. It asks who the finished form is actually for, and it proposes an answer I had not considered: not the patient, and not me, but the claims examiner in another state who will give the document 90 seconds and a decision. I had been building a tool to get paperwork off my desk. The question reframes the project as a tool that helps a stranger say yes.
I sat with that question for a while. The model had refused to draft anything, and the refusal was the most useful thing an AI did for me that week.
140 lines, no code
The interviewer is 140 lines of markdown. It contains no code, calls no API, and requires nothing beyond the file itself. The logic runs in three moves.
The first move is triage. The skill decides whether the task deserves an interview at all. A quick email gets none. A medium task gets two or three questions in a single round. A complex, multi-section deliverable gets four to seven, sometimes across two rounds.
The second move is a context scan. Before asking me anything, the model reads what it already has: the conversation, its memory of prior work, my notes database. It drafts a working spec from that material and marks the genuine gaps, the decisions it cannot make on my behalf. The instruction in the file is explicit: do not ask what you already know.
The third move is the interview. Every question must pass a single test written into the file: if this question were removed, would the output be noticeably worse? Questions about tone and formatting fail that test. The surviving questions are about substance: the one thing the reader should take away, the strongest objection and how to handle it, what someone should do after engaging with the finished thing. My answers become a brief, and nothing gets built until I approve the brief.
The file states its own reason for existing better than I can. When you ask an AI to build something, it makes dozens of decisions you never specified, and most of them get filled with generic defaults. The gap between pretty good and exactly right is usually three to seven unasked questions about substance, not style.
The opposite of the usual pitch
Every other AI tool in my working day sells relief. The ambient scribe listens so I do not type. The retrieval tool reads so I do not search. The drafting assistant composes so I do not stare at a blank message. The whole category is pitched on subtracted friction, and most days I am glad to have it.
The interviewer adds friction on purpose, at the exact moment I want relief, and the friction is the product. We have learned to judge AI by the quality of its answers. This skill earns its place by the quality of the question it makes me answer first. At the level of the transaction, it cost me 11 minutes. At the level of the deliverable, it was the fastest thing in the room, because the spec I approved that night has not needed a second version.
The oldest move in medicine
Physicians should recognize this pattern, because it is ours. A patient does not arrive with a spec. A patient arrives with a chief complaint, which is a request the way “write me a handout” is a request: true, urgent, and radically underspecified. “Chest pain” is not an order for a troponin. It is the opening line of an interview.
The evidence for that interview is older than most of the technology in my exam room. Hampton and colleagues published the classic study in the British Medical Journal in 1975. After reading the referral letter and taking the history, before laying a hand on the patient and before a single test, physicians had already reached the final diagnosis in 66 of 80 new medical outpatients. The physical examination changed the answer in seven patients. Laboratory investigation changed it in another seven. The interview was the diagnostic instrument. Fifty-one years later, the same discipline applies when the presenting complaint is mine and the consultant is a language model.
The interviewing pattern has earned its keep in four places in my clinical work so far.
The first is the patient handout. Ask the questions before drafting and the handout changes: the one thing this patient should retain, the worry underneath their question, the action they should take next, the reading level all of that requires. The generic version of a continuous glucose monitor handout explains the sensor. The interviewed version answers the fear.
The second is the workflow memo. A rooming-protocol change or a huddle script lives or dies on whose behavior has to change and which objection will kill it in week two. Those are exactly the questions a good brief forces and exactly the ones a generic draft skips.
The third is the talk. A noon-conference outline built from “make me slides on ambient scribes” is a Wikipedia entry with a template. The same outline built after answering “what should a skeptical internist do differently on Thursday morning” is a talk.
The fourth is the murky case, and this one surprised me. The pattern works with no deliverable at all. For a 64-year-old composite with four months of fatigue, 4 kilograms of weight loss, and a clean first-pass workup, I sometimes ask the model to interview me rather than answer me: what has actually been excluded versus assumed, which finding would change management, what the patient wants from the next six months. The model makes no diagnosis. It makes me articulate the case the way I would to a sharp consultant on the phone, and the articulation is frequently the answer. AI supports, AI does not decide. This is what support looks like when the cognition that needs the help is mine. Case details are de-identified before any of this, per the note at the close.
43 words, any model
I run this in Claude because skills are a native feature there and because the file can check my notes before it asks me anything. None of the value lives in the feature. The portable version is 43 words, and it pastes into the top of a chat with any frontier model, ChatGPT and Gemini included:
Before you draft anything, interview me first. Ask the three to seven questions whose answers would most change the output, propose your best-guess answer to each, and write a one-paragraph brief from my responses. Do not start building until I approve the brief.
That is the whole technology. The model brand matters less than the order of operations.
If you are a clinician working with AI tools, the question is no longer whether the model can write the handout, the memo, or the talk. It can, and it will, filling every space you left blank with a plausible default. The question is whether you decided what the thing is for before the model decided for you. An AI that interrogates your thinking before it helps is a different category of tool from an AI that answers, and right now it is the underrated one.
The disability spec is finished. Every output it produces leads with the functional language an examiner needs in the 90 seconds the file will get. I typed two sentences that Tuesday expecting a draft, and the most valuable thing I received was a question I could not answer. The draft waited until I could.
Doug Fullington, MD is a practicing internist with over 25 years in primary care. He has no financial relationships with any of the clinical AI tools discussed. He writes about AI in primary care at AI from the Exam Room (https://dfullington.substack.com/). The views expressed are his own.
A note on PHI and AI clinical tools: Even when a platform has a signed BAA, the HIPAA minimum necessary standard still applies. Most clinical questions can be answered with de-identified details (age, sex, relevant history) without names, dates of birth, or MRNs. Check your institution’s policies, which may add restrictions beyond HIPAA. Any patient descriptions are fictional examples and no PHI has been included.