Common workflows

Use these recipes to give Mizura enough context to do useful work and make the result easier to check. Each one starts with a prompt you can adapt; the agent may ask for clarification when a required choice is missing.

Explore a dataset before analysis

1. Load the file and inspect the preview, types, missing counts, and summary statistics.
2. Ask for a profile and name any columns or coding rules you already know.
3. Check the profile notes against the source documentation before choosing a method.

Profile this dataset before analysis. For each column, report its type, missingness, range or common values, and any suspicious codes or inconsistent labels. Do not recode anything yet. Tell me which issues need my decision.

The first rows are only a preview. The agent's profile looks for data-quality clues, but these are prompts for inspection, not proof that a value is missing or invalid. For large streamed files, the profile is based on a repeatable sample; row counts and some counts are exact. See Your data.

Compare two groups

1. Identify the outcome column and the column that defines the groups.
2. State the population, the comparison you care about, and any paired or clustered design.
3. Ask for a plan if you want to review the method before it runs.
4. Inspect the method card, assumptions, diagnostics, effect estimate, and uncertainty in the outputs.

Compare the mean of `outcome` between the groups in `group`. Show group sizes, missingness, a plot, the difference with a confidence interval, and an appropriate test. Check whether the design is paired and whether the method's assumptions are supported. Do not infer causation.

Replace the example column names with exact names from your file. If observations are paired, clustered, weighted, or otherwise dependent, say so; a simple two-group comparison may not fit that design. A methodology card guides the analysis but cannot validate the study design. See Methods.

Fit and check a regression

1. Name the outcome and candidate predictors, including which are categorical.
2. Describe the unit of observation and the population or time period.
3. Ask for a plan, then review the method and diagnostics before pressing Go.
4. Read the code, tables, figures, and limitations alongside the written interpretation.

Estimate the association between `sales` and `tv`, `radio`, and `newspaper` spend. Explain the target population and the observational nature of the data. Check model assumptions, influential observations, and multicollinearity; report coefficients with confidence intervals and diagnostic plots. Do not describe the estimates as causal.

Change the example to match your research design. A regression output does not establish causality, and a diagnostic plot cannot repair a weak design or unmeasured confounding. If a conclusion depends on an assumption, ask for a sensitivity analysis and review the code that produced it.

Prepare a revised dataset

1. Tell the agent the rule to apply and the reason, including how missing and special codes should be treated.
2. Ask it to preview the changes and report affected rows and columns.
3. Review the dataset-change summary before using the new version.
4. Keep the original available; the new version is saved separately and can be undone from dataset history.

Create a new dataset version that excludes rows where `status` is 'test'. Keep every other column unchanged. Report the number of rows removed, confirm the original is untouched, and show the resulting row count. Do not treat blank strings or 999 as missing unless the codebook confirms that meaning.

Dataset edits are saved as new versions; the original is not modified. The agent's change summary describes what changed, but the core does not require a separate approval before switching the session to the saved version. Inspect the version and use dataset undo if the change is wrong. See Your data.

Research a method and its limitations

1. State the method, population or domain, and the question the literature should answer.
2. Ask for a literature search and for the relevant sources to be opened.
3. Inspect Sources and separate full-text findings from abstract-only summaries.
4. Use a known DOI to trace references or later citing work when a result needs context.

Find published guidance on handling missing covariate data in observational regression. Open the relevant sources, summarize the conditions each recommendation assumes, and identify disagreements or limitations. Do not infer findings from search snippets.

Success means the summary links to retrieved material and states retrieval limits. A formatted citation alone does not establish that the article was read. See Follow citation chains and Evidence & literature.

Write a report from checked results

1. Complete the analyses and inspect the outputs, diagnostics, and source material.
2. Ask for a report that names the research question, audience, language, and reporting style.
3. Review the draft and generated PDF or other formats; revise sections as needed.
4. Export a notebook, Quarto document, or replication package when you need an analysis record.

Write a concise report in English for an academic audience using APA 7. Answer the research question from this session. Use only results from executed analyses, include the relevant diagnostics and limitations, and cite only sources opened in this session. Do not add unsupported numbers or claim causation.

Reports can be rejected when numbers are not grounded in executed results, citations were not opened, or required results are missing. Generated reports and exports still need your review. A replication package can contain the session's dataset and outputs; inspect it before sharing. See Reports & exports.