Plan it before you run it.
A guided planning tool for healthcare and clinical-research students. An idea becomes a research question, a study design, a statistical plan you confirm, and the R to run it.
Your results, before you have them
95% CIMost student projects fail at the planning stage, months before anyone touches the data.
The question is too broad to answer. The design cannot support the sentence the student wants to write at the end. The sample size was worked backwards from whatever records happened to exist. Draft asks the questions a methodologist would ask, in the order they would ask them, and tells you what follows from each answer.
What your design cannot tell you
Twenty-eight designs, reached by how people enter your study rather than by what sounds strongest. Each arrives with its reporting checklist and the sentences it does and does not support.
Code that matches the plan you confirmed
Written with your own column names, against practice data so you can debug the whole thing while your data access is still being approved. The script states the rule, runs the check, and prints which branch it took.
# PRE-DECIDED RULE — written before the data were seen. # Default: Welch's t-test If it fails: Mann-Whitney U normal_ok <- all(sw$p_shapiro >= 0.05) if (normal_ok) { cat("DECISION: both groups pass the check.\n") res <- t.test(los_days ~ ward, data = dat) } else { cat("DECISION: fallback runs: Mann-Whitney.\n") res <- wilcox.test(los_days ~ ward, data = dat) }
Three decisions, in an order that cannot be skipped.
Each narrows the next. Change any answer and everything downstream updates — including the code, which hides itself again until you re-confirm.
Settle the question
Population, exposure or denominator, outcome, timepoint. Six starting points, from “I have an idea” to “my proposal is written and I need the design”.
Confirm the plan
Every comparison written as a default test, the assumption it rests on, how you check it, and what runs instead if the check fails. You confirm before any code exists.
Run the analysis
R built from the confirmed plan, the figure that matches your test, and a protocol or proposal draft to take to supervision.
Everything it does, grouped by when you need it.
Nothing here is a chatbot answering from memory. Each panel is built from the answers you have already given, and says what it is assuming.
Before you have a question
Question builder
Turns a topic into a population, an exposure or denominator, an outcome and a timepoint — and tells you which part is still missing.
Six starting routes
An idea, a question, a design to check, an analysis to plan, a protocol to write, a proposal to write. You only see the sections your route needs.
The ? beside every question
What is being asked, why it matters, a strong answer and a weak one. Works offline, with no call to anything.
Feasibility and access
Who holds the data, what approvals it needs, and how long each takes — before the design depends on getting them.
Choosing the design
28 study designs
Reached by how people enter your study and when things were measured, not by what sounds strongest. Descriptive, observational, experimental and quasi-experimental.
What it cannot tell you
Every design arrives with the sentences it supports, the sentences it does not, and the reason — which is what a limitations section is made of.
Reporting checklists
CONSORT and SPIRIT for trials, STROBE for observational work, plus RECORD, TREND and SQUIRE where they apply.
Bias, named early
What a reviewer will raise first, matched to your design, while you can still change the design rather than apologise for it.
Planning the analysis
Sample size, both directions
Two proportions, two means, one proportion, a correlation — or, when the number is fixed by what exists, what that number can actually detect.
The statistical plan
Default test, assumption, how you check it, what runs instead. Confirmed before any code appears, and it lapses if you change an answer behind it.
Missing data
How much, where, and why — because what you are allowed to do next depends entirely on the last of those.
Matched and paired designs
Conditional logistic for matched sets, paired tests for repeated measures, and a stop panel when the answers say both independent and matched.
Running it
R, written to your columns
Eleven blocks, each with what you should see, two ways it goes wrong, and a line to try yourself. A first block on installing R if you have never opened it.
Practice data
Generated with your own column names, so you can debug the whole script while your access is still being approved. Nobody debugs code against patient data.
Error translator
Eighteen common R errors matched by pattern, each with a plain cause and a fix. Fixed list, no AI, works offline.
The figure that fits
Thirteen chart types scored against your test — fits, works, or poor fit, each with the reason. Plus ggpubr, ggstatsplot, tidyplots and ggalluvial, taught rather than listed.
Writing it up
Protocol and proposal drafts
Structured to SPIRIT or to the items in your reporting checklist, filled from what you have answered, with the gaps marked rather than invented.
Supervision handover
A short document listing what is decided, what is still open, and the questions worth asking — so the meeting is spent on the gaps.
Your own reference library
Upload your course handbook or local guidance and ask questions of it. Answers quote the passage, and say so when the documents do not cover it.
Edit every question
Reword anything, change when it appears, add your own. The content is data, not code, so it is yours to adapt to your course.
No student has used this yet. Here is what has been verified instead.
A methods tool that cannot show its working is asking to be trusted on manner alone. Every branch of the generated R has been run, and the numbers compared against published analyses of real studies.
| Study | Design Draft named | Published | Draft’s code |
|---|---|---|---|
| Adjuvant chemotherapy in colon cancer | Parallel-group RCT, survival | HR 0.599 | HR 0.599 |
| Risk factors for low birth weight | Cross-sectional, yes/no outcome | χ² 4.92, p 0.026 | χ² 4.92, p 0.026 |
| Progabide for partial seizures | RCT, count outcome | dispersion 10.9 | switched to neg. binomial |
| Anorexia treatment, three arms | RCT, number with a before-value | +8.66 lb | +8.66 lb |
| Infertility after abortion | Matched case-control | OR 7.29 | OR 7.29 |
The fifth study is the one worth knowing about. Draft got it wrong at first — it sent a matched case-control to the wrong test, which would have reported less than half the real effect with a small p-value and nothing to warn you. That is what the exercise was for, and it is fixed. The full record of every defect is kept with the project rather than hidden.
You can. It will answer every question, including the ones you should not have asked yet.
A general assistant is good at explaining a method you name. It is not built to stop you naming the wrong one, and it has no memory of the denominator you settled on four questions ago.
Asking a general assistant
Capable, fast, and happy to agree with you
- × Answers whatever you ask, including a t-test for a design that cannot support one
- × Each answer starts fresh; nothing checks it against what you decided earlier
- × Gives you code immediately, before the plan exists to judge it against
- × Will produce a reporting checklist, but not notice that it is the wrong one
- × Answers from everything it has read, so you cannot check where an answer came from
- × Leaves no record of what you decided or when
Draft
Narrower, slower, and willing to tell you no
- ✓ Routes to a design from how people enter your study, then says what it cannot show
- ✓ Every panel is built from your earlier answers, and flags it when two of them conflict
- ✓ No code until you have read and confirmed the statistical plan
- ✓ The checklist follows from the design, with its extensions
- ✓ Library answers quote the passage, and refuse when your documents do not cover it
- ✓ Produces a dated plan, a protocol draft and a handover you can take to supervision
What Draft will not do.
This list is on the tool’s own screens too. A planning tool that oversells itself is worse than no planning tool, because a student will believe it.
It is not a substitute for any of these.
- A supervisor. It can tell you which questions to bring to the meeting. It cannot have the meeting.
- A statistician. Every plan it writes says it needs sign-off, because choosing an analysis is a clinical judgement and not an arithmetic one.
- Ethics approval or governance. It will tell you which approvals your design is likely to need. Your institution decides, not this.
- A submittable protocol. What it produces is a structured draft with the gaps marked, meant to be argued with.
- Anywhere to put patient data. No names, no dates of birth, no record numbers, no real datasets. It is for planning a study, not running one.
- An authority on your own course. Upload your handbook and it answers from that. Without it, it says the documents do not cover your question — which sends you to a person, correctly.
The ones worth asking first.
What does it cost, and how do I get in?
It is free. There is no trial and no card.
Access is by invitation while it is being piloted. Your email address is added to an approved list; when you open it you are sent a sign-in code, and after that a second step on your phone or laptop confirms it is you. If your address is not on the list, the page will not let you through.
How do I get added to the list?
Ask. It is deliberately a short list during the pilot — the tool has been tested hard against published data and not yet against real students, and the first few people using it are people whose feedback can be followed up properly.
Write to [ your contact address ] with the course or project you would use it for.
Where do my answers go?
Signing in controls who gets through the door. It does not change what happens inside: your answers stay in your own browser as you work, and in the file you choose to save. Nothing you type into Draft is stored on a server, which also means nothing follows you to another computer — use Save a copy before you close the tab, and open that file to carry on elsewhere.
Because of this, the one thing worth checking before a class is that saving works on your machines. The tool has a built-in check for exactly that.
Can I put my study data into it?
No, and the tool says so on every screen. Draft is for planning a study, not running one. No names, dates of birth, record numbers, clinical notes or real participant datasets — not in the questions, and not in the reference library.
The R it writes comes with generated practice data using your own column names, so you can debug the whole script before you have access to anything real.
Do I need to know R?
No. Tell it you have never opened R and the script gains a first block covering
installing R and RStudio, what the console is, how to run a line, and 2 + 2.
If your course requires SPSS or Stata, follow your course. The order of the steps is the same in every package: check the data, count what is missing, describe the groups, then analyse.
Why does it hide the code until I confirm the plan?
Because code is easy to produce and hard to defend, and a student who gets the script first will work backwards from it. The plan states what runs by default, the assumption it rests on, how that assumption is checked, and what runs instead when the check fails. Confirming it records that you decided before you saw the data.
Change an answer the plan was built from and the confirmation lapses. The code hides again until you have read the new plan.
Is the statistics right?
Every branch has been executed rather than read. Sample sizes match the
pwr package within one to three people, erring upwards. Prevalence and
incidence intervals match epitools to several decimal places. Five published
studies were planned in Draft and their published results reproduced by the code it wrote.
That establishes the tool computes what it claims to. It does not establish that the plan is right for your study, which is why every plan it writes says it needs sign-off.
Can I adapt it for my own course?
Yes. Every question can be reworded, moved, hidden behind a condition, or replaced, and you can add your own. The content is data rather than code. You can also upload your department’s handbook and local guidance so the Ask box answers from your documents instead of in general.
Is it finished?
No. It is a working prototype that has been tested hard against real published data and not yet against real students. That pilot is the next thing, and the feedback question that matters is “what did you change about your study because of this?” — where a blank answer is a real finding.
Open it with a study you actually have.
About forty minutes end to end, and you leave with a plan, a script, and something to put in front of a supervisor.