A good prompt names the report, the marker, the unit, and the date, and asks for a value you can check on the page. A bad prompt asks the assistant to judge you. Each of the ten prompts below produces an answer you can verify against a specific line in a specific report.
The rule behind every prompt
Three habits separate a useful answer from a plausible one.
- Name the source. "On my report dated 12 March" beats "in my results". If the assistant cannot find that report, you want it to say so.
- Ask for the value with its unit and range. A number without a unit is not a result, and a number without a range cannot be read.
- Ask for a checkable claim. Every sentence should point at something printed on a report, not at a general fact about a disease.
Ten prompts that produce checkable answers
| Prompt | A good answer gives you | How to check it |
|---|---|---|
| "List every result on my report dated [date] that is outside its printed reference range, with the value, unit, and range." | A list that matches the report's own flags, in the same units. | Compare against the flagged rows on the report. |
| "Show my [marker] values in date order, with units, and tell me which report each came from." | One row per report, with a date and a link or a report name. | Open the earliest and the latest report. |
| "Which of my markers changed between [date] and [date], and by how much in each one's own unit?" | The two values, the difference, and the unit of each marker. | Check that both values are on the reports you named. |
| "Summarise the report dated [date] in five bullet points, quoting the exact values." | Five points, each quoting a number from that page. | Read one quote against the page. |
| "Which results on my [date] report have no reference range, and what does the report say about them?" | The tests without a usable range, and any note printed beside them. | Look at the range column on those rows. |
| "Does any report mention haemolysis, lipemia, dilution, or a repeated test? Quote the remark." | The exact remark and the report it appears on. | Find the remark in the report's notes. |
| "For [marker], what are the earliest and latest values you can see, and what are their dates?" | Two values, two dates, and the direction of travel. | Open both reports at that marker. |
| "List the panels on my [date] report, and which tests belong to each panel." | The report's own grouping, not a textbook one. | Compare with the section headings on the report. |
| "Which of my reports note whether the sample was fasting or non-fasting?" | The reports that state it, and what they state. | Check the fasting note or flag. |
| "Draft five questions for my doctor, based only on the values in my reports, and cite the value behind each question." | Five questions, each with the number that prompted it. | Confirm every cited number appears on a report. |
What a good answer looks like
- It quotes values with units, and it uses the ranges printed on the reports rather than a generic chart.
- It says which report a number came from, or links to it.
- It separates what the data shows from what might explain it.
- It says plainly when something is not in the records, instead of filling the gap.
- It does not turn one reading into a diagnosis.
Prompts that produce confident nonsense
Some questions have no checkable answer, so the model produces the most plausible-sounding one. These are the prompts to avoid, whether you are in a chat window or in DoctanoAI.
| Prompt | Why it goes wrong | Ask instead |
|---|---|---|
| "Do I have [condition]?" | A diagnosis needs a clinician and usually more than one test. | "Which of my results are outside their range, and which ones relate to [system]?" |
| "Is 5.9 bad?" | The number has no marker, no unit, and no range. | "What is my HbA1c, its unit, its range, and how has it moved across my reports?" |
| "What dose should I take?" | Prescribing is a clinical act. DoctanoAI does not recommend doses. | "What should I ask my doctor about this result?" |
| "Am I healthy?" | No set of lab results can answer that. | "Summarise the results I have, and list what is outside its range." |
| "Which supplement fixes this?" | It invites a treatment recommendation the data cannot support. | "Which of my markers are low, and which reports show it?" |
| "Should I stop my medicine?" | Nobody should change a prescription on the strength of a chat reply. | "Which of my results changed after [date]?" and then ask the prescriber. |
Checking an answer inside Doctano
Doctano keeps every extracted value linked to the report it came from. Open the report from a result, then use View original to compare the extracted value with the uploaded file. If a number in a chat reply does not match the page, the page wins.
DoctanoAI can use the selected profile's values, history, out-of-range results, and report text, and it may point back to the relevant reports. It cannot reliably answer from information that is not in the profile, so a question about a report you never uploaded should get an honest "not in your records" rather than a guess.
What not to expect
- An assistant cannot see a report you did not give it, or a conversation that has ended.
- It cannot confirm that a lab's extraction or your transcription is correct. Compare with the source.
- It cannot weigh symptoms, an examination, or a family history the way a clinician can.
- It must not be used for an emergency. Contact a doctor or local emergency services.