Drafting a medical research plan with AI: inputs, outputs and tool choice

What materials should an actual task receive?

This is a handoff checklist, not evidence that the teaching scenario already has these materials.

What reviewable outputs look like: teaching excerpts only

1. Research-question statement

   Propose examining the association between preoperative recorded smoking status and readmission within 30 days in adults undergoing elective surgery. Surgery type, record period and available sample remain unconfirmed. Do not turn association into a causal claim.

2. Variable-definition table

   Exposure: recorded preoperative smoking status; categories and timing require verification. Proposed index: surgery date, checked against actual records. Outcome: readmission within 30 days; establish whether other-hospital records are captured. An absent record is not automatically a non-event.

3. Plan paragraph

   Propose a retrospective observational cohort. Specify eligibility, baseline records, outcome ascertainment and follow-up coverage before discussing confounding, missingness and sensitivity analyses. Do not freeze a model or run statistics before ethics, authorization, event counts and variable availability are confirmed.

4. Gaps and next step

   Only the idea has been supplied; fields, literature and authorization need checking. Add materials and review the design next, rather than generating a paper that presents the study as completed. STROBE checks reporting content; it does not validate this plan.

Choosing DeepSeek, ChatGPT, Codex, WorkBuddy or RigorMed

Choose by the work to hand over. These are workflow suggestions, not a matched performance test or a partnership or integration claim.

Five stages: tool assistance and researcher checks

One tool may cover several stages, but producing an answer does not justify skipping source and method checks. This is a responsibility checklist, not certification of any product's capabilities.

PubMed principally supplies biomedical citations and abstracts and may link to full text. Finding a record does not mean its full text has been read or its quality assessed. Applicable STROBE checklists help check observational reporting, but do not replace sound design.

Common AI tool searches: start with the research task and its checks

These groups address tool names researchers ask about and provide practical checks, not rankings, feature certifications or an integration list. Consult official documentation for versions, plans and access conditions. Adding 'research' to a tool name does not mean a study is complete.

This table provides RigorMed task-checking recommendations, not promises about any tool's current features. Naming a brand does not imply partnership, official affiliation, integration or superior quality.

Tool queries lead to this shared workflow guide. PubMed strategies, data cleaning, Table 1 and submission checks remain separate topic guides. The scientific figures page explains chart and graphical abstract limits; synonyms do not justify duplicate thin pages.

Adaptable research prompts and verification steps

1. Literature notes: use only the selected papers I am entitled to use. For each paper, list its question, design, population, exposure or intervention, outcomes, main findings and limitations. Give a source and passage for every item; say 'not reported' when absent. Afterwards, check each original paper.
2. Review and references: outline a review using verified papers, separating agreement, conflicts and gaps. Add no new citations and do not turn association into causation. Verify PMID/DOI and the original paper before using the outline; it is not a formal systematic review.
3. Code and cleaning: use the approved data dictionary and rules to generate replayable R/Python processing for public or synthetic examples. Record each join, exclusion and sample count. Do not automatically impute missing values, select variables or change the analysis plan. Run the code and inspect actual outputs before continuing.
4. Manuscript and figures: use only executed analysis outputs to organize Table 1, figure descriptions and manuscript text. List inconsistencies in numbers, units, citations and limitations. Without outputs, retain the gaps; do not invent significance, effect estimates or promises of submission readiness.

Current RigorMed entry points: hand over materials before proceeding

1. Choose a task, not a capability certificate

   The public site offers task entry points for different disciplines. In the workspace, describe the research objective, available materials and intended deliverable. This guide still concerns medical research: its medical designs or survey methods cannot simply be transferred to another discipline. Selecting an entry point does not mean a task has run or a human expert has reviewed it.

2. Separate bibliographic search from full-text reading

   The literature entry finds records and available abstracts; asking a model to read a paper is a separate step. Record sources, search terms and dates, then verify lawful access to the full text. With only a title or abstract, do not ask the tool to invent body details, figure values or conclusions the authors did not report.

3. Use a case as a reference, not a new result

   The case directory covers abstract editing, literature translation, PDF reading, research planning, data charts and paper presentations. Inspect the original materials, output language and limitations before defining a task with materials you are entitled to use. A case output is not a new task result. Interface language does not imply translated deliverables; illustrative data are not real research evidence.

4. Check requested materials and billing before execution

   At task entry or when more information is requested, complete the current options, custom input or attachments. State missing, unreadable or unauthorized materials explicitly instead of asking a model to guess. Check the current model selector, credit multiplier and billing explanation before confirming execution. A selectable model, submitted form or credit charge does not establish correct sources, completed calculations or an acceptable deliverable.

Task handoffs illustrated by a public-data question

1. Specify the question instead of requesting a conclusion

   As a conceptual example using a public NHANES cycle: is physical activity associated cross-sectionally with measured blood pressure among US adults? A real study must specify and verify cycles, population, exposure, outcome, and the association to report. This is not an individual follow-up or causal-effect question.

2. Check published evidence

   Draft a PubMed query using physical activity, blood pressure, and population concepts. Inspect Search Details and test known relevant publications. Retain dates, queries, and identifiers. AI summaries are reading aids, not substitutes for original papers and design checks.

3. Confirm the data can answer the question

   Check the CDC/NCHS variable documentation for the chosen cycle: physical activity questionnaires, blood pressure measurements, eligibility, special codes, and design variables. For several cycles, also assess measurement and definition comparability. Do not allow automatic merging based on column names.

4. Freeze cleaning and statistical rules before analysis

   Preserve original files. Record the dictionary, missingness reasons, keys, and transformations before selecting estimands, potential confounders, weights, strata, and PSUs. Tools may execute confirmed rules; do not change models or exclusions retrospectively because results are significant.

5. Check results and deliver reviewable materials

   If formal analysis is authorized, check sample flow, effect units, intervals, and agreement between tables, figures, and text. Explain cross-sectional limitations. No data were analyzed here, no estimates were produced, and this illustration does not claim RigorMed completed the study for a user. Shareable reproducibility materials remain subject to source conditions.

Seven checks before the next stage

Regardless of the tool, a reviewable handoff should answer at least these questions:

• Are the question, population, temporal structure, and permissible conclusions explicit?
• Can cited papers, registry records, and documentation be traced to original sources and versions?
• Are strategies, dates, screening criteria, and uncovered sources retained?
• Have use conditions, dictionaries, missing codes, and sample flow been checked?
• Were methods determined before examining results and do they fit the cohort or sampling design?
• Does every figure, table, and numerical statement map to actual output rather than AI-generated additions?
• Are limitations, ethics and author responsibilities, sharing scope, and reproducibility conditions stated?

If evidence is missing, mark it pending and remain at the current stage. Do not describe drafts, simulated reviews, or workflow illustrations as completed research or clinical advice.

Frequently asked questions