Best Practices
Keep each study separate, keep the intermediate results, and check anything you will rely on.
Organize by Study
One project per study
Keep the data, papers, conversations, notes, and results for one study together. Use separate projects for unrelated clients, protocols, or groups of people.
Name projects so you can find them later
Prefer “TRIAL-042 primary outcome · 2026-07” over “New Project.” Use labels for stage, topic, or owner, and never for identifying a participant.
Preserve Raw Data
- Keep source files unchanged, and back them up outside CATO as your institution requires.
- Upload the data dictionary, protocol, analysis plan, and codebook with the data.
- Save cleaned data as a new file rather than replacing the source.
- Record units, missing-value codes, reference build, time zone, and reference levels.
- Check the expected row count after every merge.
Be Specific About Your Data
Mention exact datasets with @, say which version you mean, and attach only the papers the question needs. If two files have similar names, say which one is right and how they differ.
Vague
“Analyze the final data and make a graph.”
Specific
“Use @analysis_ready_v3.parquet, version 4. Plot the prespecified adjusted effect with 95% CI and save the plotting data and editable figure.”
Plan Before Important Work
Ask CATO to plan first when the work includes exclusions, imputation, model selection, multiple endpoints, eligibility decisions, or anything going into a publication. Read the plan for assumptions, and ask for changes before you tell it to go ahead.
A good plan names the inputs and versions, transformations, method, checks, expected outputs, when to stop, and what CATO should ask you rather than assume.
Choose the Right Mode
- Standard: routine analysis and writing, exploring data, summaries, formatting, and small edits.
- Research: harder methods and deeper evidence work.
A clearer request usually helps more than switching mode. Start with Standard, and move to Research when the problem is hard rather than unclear.
One Conversation per Objective
Start a new conversation when you move from exploring to a prespecified analysis, from one manuscript section to another, or to an unrelated question. CATO then has less unrelated context to work around, and the work is easier to trace later.
Handling Long Runs
- If a response is reconnecting, wait for it before sending the request again.
- Press Stop once if the inputs or the objective are wrong, and give it a moment to finish stopping.
- After an interruption, check what was finished before you ask again.
- Name retries clearly, so their outputs are not confused with the original run.
Treat the Files as the Record
Open output files, notebooks, figures, and manuscripts in the Workspace, check their Provenance, and use the exact version you mean in later work. A chat summary helps you find things, but the files and their sources are what you can check.
Make Notebooks Re-runnable
- Ask for imports, settings, seeds, and data loading in the first cells.
- Avoid notebooks that only work when cells run in one particular order.
- Save important intermediate data and parameters as files.
- Ask CATO for a clean run from the top before you finish.
- Download the code and package versions someone would need to reproduce it.
Check Figures Carefully
- Keep the plotting data, code, and original figure version.
- Use versions to keep analysis changes separate from visual edits.
- Check axes, transformations, denominators, error bars, legends, clipping, and annotations.
- Use accessible colors, and look at the figure at its final publication size.
- After editing, check the downloaded PNG, SVG, or PDF, not only the preview.
Keep Literature Work Reproducible
- Save the question, date, year range, and any warnings.
- Write the question as a full sentence, then read down the results. On most accounts nothing is screened out for you. On an organization account with restricted sources, only the papers screened in are shown.
- Look for duplicate versions, corrections, and retractions.
- Check extracted evidence and eligibility against the full text where you have it.
- Use watches to stay current, but allow for indexing delays.
- When sending papers to Chat, say which sources to use and ask for support for each claim.
Treat Notes as Reminders, Not Findings
CATO keeps notes on a project in its Memory tab, and notes about how you work in Settings → Memory. Look through them from time to time, and edit or delete any that are wrong or out of date.
A note is not evidence. Results belong in cited, versioned files whose inputs and methods can be checked.
Check Before You Share
| Inputs | The right project, files, versions, cohorts, variables, units, and identifiers |
| Transformations | Exclusions, missing-data handling, recoding, normalization, joins, and derived variables |
| Method | The quantity being estimated, model, assumptions, multiple testing, leakage, convergence, and sensitivity |
| Outputs | Expected rows, effect sizes, uncertainty, diagnostics, figure marks, labels, and the downloaded file |
| Evidence | The right paper, available full text, supporting passage, exact quotations, and retractions or corrections |
| Reproduction | Saved code, seed, package versions, logs, source versions, and a clean re-run that matches |
Share Safely
- Prefer a checked file to a screenshot, which leaves out where a result came from.
- Before you make a public conversation or figure link, confirm it shows no restricted data.
- Take public links down when they are no longer needed.
- Do not put passwords, keys, protected health information, or confidential raw data in support tickets.
- Follow your institution's rules on data use, authorship, disclosure, and retention.
Keep Your Own Research Record
CATO stores your files and their versions, but it does not replace your own record keeping. Keep final data, code, notebooks, searches, evidence tables, manuscripts, and figures in the repository your institution, protocol, funder, or regulator requires.