Choose one of three approaches. The steps for that option appear below.
Choose an option above. On a phone, swipe the tabs left or right to see all options →
Claude drafts the test plan and tester instructions from your case details. You check every line and sign off.
Write case details
Draft test plan
Read against case details
Draft tester sheet
Check, sign off
Use: The clinician
Write down in plain text: intended use, approved cases labelled like CASE-01 with the finding you consider correct, your pass and fail rules.
Check every sample case is approved for this test under local rules, and names, dates of birth, record numbers, scan files and anything identifying are removed before any pasting.
Use: Claude
Ask Claude for a test plan draft with a scoring sheet layout and open questions, using only what you paste.
What to paste into the tool
Task: A draft test plan and a blank scoring sheet layout for checking a clinical AI tool against sample cases. Use only the packet below. Do not add cases, findings, accuracy figures or thresholds that are not in the packet. Write: purpose, what is in and out of scope, one row per case with the clinician's stated correct finding, the columns a tester fills in (tool output, match yes/no/partly, tester note), the pass and fail rules exactly as I wrote them, and a list of questions where my packet is unclear. State at the top that this plan does not permit use on patients or any clinical decision. Do not diagnose anything. Packet: [paste the test packet].
Use: The clinician
Read the draft against your case details line by line, delete anything you did not write, answer the questions, add missing case types.
Check no invented case, finding, number or threshold remains, the pass and fail rules are your own, and the plan is marked test only, not for patient use.
Use: Claude
Ask Claude for a one-page tester instruction sheet from the corrected plan, adding no new rules.
What to paste into the tool
Task: A one-page instruction sheet for a person who will record results in the scoring sheet. Use only the plan below. Say how to record each result, what to do when the tool output is unclear (write 'unclear' and stop, do not guess), and that nobody uses the tool on a patient or acts on its output. Add no new rules. Plan: [paste the corrected test plan].
Use: The clinician
Read the sheet against the plan, remove added rules, sign off the final plan yourself, and store it where the test team can read it.
The clinician, not the assistant, approves the plan and decides whether and when a test is run.
Elicit finds published papers on testing diagnostic AI tools. You pick design points, then Claude drafts the plan.
Write search question
Find published methods
Choose design points
Draft around points
Check, sign off
Use: The clinician
Write the question for the literature search in plain words, for example 'how have image-based diagnostic AI tools been confirmed on sample cases'.
The clinician checks the question holds no case, patient or organisation details. Do not include any case details or patient data.
Use: Elicit
Ask Elicit for a table of papers on testing diagnostic AI tools, with design, sample and stated limits.
What to paste into the tool
Task: A table of published studies on testing diagnostic AI tools on sample cases. Question: [your search question]. For each paper give citation, test design, how cases were chosen, how correct answers were set, and limits stated by the authors. Quote the sentence each cell comes from. Say 'not stated' where a paper does not say.
Use: The clinician
Open every cited paper and confirm each cell; keep only real test designs; write three to six design points for your plan.
The clinician checks each design point comes from a paper the clinician opened and read, with its citation.
Use: Claude
Ask Claude for a test plan draft that names which design point each part follows, using your case list and rules.
What to paste into the tool
Task: A draft test plan for checking a clinical AI tool against sample cases. Use only the design points and the case list below. For each part of the plan name which design point it follows. Do not add cases, findings, figures or thresholds. List open questions. State at the top that this plan does not permit use on patients or any clinical decision. Design points: [paste]. Cases and pass and fail rules: [paste list].
Use: The clinician
Check each plan part against its named design point and your case list, remove anything invented, sign off, keep the citations with the plan.
The clinician approves the plan. No test is run until a qualified person decides.
ChatGPT drafts the plan from your checked material, then lists gaps as questions only. You check and sign off.
Ready cases, rules
Draft plan, list gaps
Check, sign off
Use: The clinician
Your approved case list and rules, checked and ready to paste.
The clinician checks the material holds nothing that could identify a person, and every rule is the clinician's own.
Use: ChatGPT
Ask ChatGPT to draft the test plan from your material, then list gaps and unclear points as questions only.
What to paste into the tool
Task: Draft a clinical AI test plan from my approved material. Cover: the purpose, the approved sample cases by label, the correct finding for each case, pass and fail rules, and how results are recorded. Use only my supplied cases and rules; do not add cases, findings or rules of your own. If something needed is missing, ask me a question instead of inventing it. This is a plan only; do not include wording that reads as permission to use the tool on patients. My approved case list and rules: [paste them here]. Task: A list of questions about gaps in the test plan below. Do not rewrite the plan or add content. List: case types that seem missing, pass and fail rules that are ambiguous, places where two testers could score the same output differently, and any wording that could be read as permission to use the tool on patients. Write each as a question. Plan: [paste the plan].
Use: The clinician
Check the draft against your approved cases and rules, go through the gap questions, change only where you agree, sign off yourself.
The clinician checks every case label, finding and rule matches the approved material, decides which questions matter and approves the plan.
Best when: You know the cases and what counts as correct, and want speed
What it does: Claude drafts plan and instructions; you check both
Strong at: Two person checks catch invented rules before sign-off
Limit: Your check is required; do not treat a draft as approval
Free plan: Yes, limited (5-hour session and weekly caps)
Best when: You want the plan to follow how others have tested similar tools
What it does: Elicit finds papers; Claude drafts around checked design points
Strong at: Design points come from papers you open and check
Limit: Opening and reading papers takes the most time
Free plan: Elicit: Yes, limited (Research Agent/Reports); Claude: Yes, limited (5-hour, weekly caps)
Best when: You want the assistant to draft while the clinician keeps every decision
What it does: ChatGPT drafts and lists gaps as questions; you check and sign
Strong at: Clinician time goes on checking and decisions
Limit: One merged drafting pass; gaps come as questions to judge
Free plan: Yes, limited (images, voice, deep research, Codex)
Before you paste into any AI tool, remove any confidential information. Put it back at the end.
Deciding between ChatGPT and Claude for this? See the comparison.
Never paste credentials or full payment details.
Every assistant step is followed by your check of the output against the clinician's own material.
A qualified clinician signs the final plan. Nothing is run on patients.
Nothing is sent, signed, approved or deployed by the assistant.
Check the result against your original inputs. Approve it before sharing or acting.