By Davie Chen, SciDraw AI team

Disclosure: The author works on SciDraw AI. All numbers in this tutorial are synthetic teaching data. Figures 1 and 3 are AI-generated editorial illustrations. Figure 2 was calculated and plotted with Python. None is a screenshot or measured output from SciDraw AI.

A researcher preparing a presentation often needs two different visuals: a chart that shows the observations and a graphical abstract that explains the study. These visuals have different jobs. A chart must preserve numerical evidence. A graphical abstract must preserve the relationship between the research question, method, result, and interpretation.

AI assistance can help with both, but a good-looking draft is only the beginning. This tutorial develops a small example, shows what to ask SciDraw AI to produce, and gives concrete checks to make before sharing the result. The same checks work whether the eventual figure is made in a browser, a spreadsheet, or a plotting script.

1. Begin with a question that the table can answer

Our example asks: how do recorded responses differ between groups A and B? We will use six synthetic observations per group, measured in arbitrary units. There is no actual experiment, biological mechanism, or clinical conclusion behind these numbers.

The rows are a convenient way to display the two lists. They do not represent matched pairs. Connecting A1 to B1 with a line would invent a relationship that this example does not contain.

The arithmetic means are 10 and 14. Those values are useful checkpoints, but they are not the whole story: group B also has a wider observed range. A figure showing only two bars would conceal much of that information.

Figure 1. Keep the source table, numerical review, and visual summary connected. AI-generated workflow illustration; its miniature charts are conceptual, not plots of the twelve observations. No product interface is depicted.

2. Prepare an uploadable table

Create a CSV with three columns: sample_id, group, and response_au. Use one observation per row, so this example has twelve rows rather than six. Assign unique identifiers such as A1 and B1. Keep numerical cells as plain numbers; put the unit in the column name or accompanying description.

Before uploading, check that the file contains twelve observations, exactly two groups, and no missing response values. The repeated value 10 in group A is valid: equal measurements do not automatically mean duplicate records. Check identifiers when looking for duplication.

For your own data, also check decimal separators, inconsistent group names, unexpected negative values, and whether zeros mean a measurement or missing information. Resolve those issues using the study documentation. Do not remove an observation simply because it makes the visual less tidy.

Retain the original file and save any cleaned version separately. If a collaborator asks why the chart changed, you should be able to trace the change to a specific row, transformation, or presentation decision.

3. Give SciVis an explicit plotting brief

Open SciDraw AI’s scientific data visualization workspace and upload the prepared CSV. The interface also accepts spreadsheet and other tabular formats, but a simple CSV makes this example easy to inspect.

A useful prompt describes the data mapping and the visual rules:

Create a dot plot of response_au by group. Use all 12 observations. Put group A on the left and group B on the right. Use small horizontal offsets to reveal overlapping dots, without changing their vertical values. Add a short horizontal line for each group mean. Label the vertical axis “Response (arbitrary units)”. Start that axis at zero. Do not add significance stars, p-values, trend lines, or a causal conclusion. Mark the chart as synthetic teaching data.

This brief gives the model fewer opportunities to make silent decisions. “Make a professional scientific chart” leaves the chart type, summary statistic, and statistical annotations unspecified.

Treat the first generated chart as a draft. If a required feature is missing, request that specific correction and check the revised output again. The prompt is a request, not a guarantee that every constraint has been followed.

4. Audit values before adjusting style

First count the points: there should be six in each group. Then check the ranges, which are 8–12 for A and 10–19 for B. Finally, check the mean markers against 10 and 14. A horizontal offset can separate overlapping observations; a vertical offset would change their apparent numerical values.

Figure 2. Both panels use the same twelve synthetic observations and the same vertical scale. The right panel makes individual values and their spread visible. This independently calculated reference plot was produced with Python, not generated by SciDraw AI.

The figure above supplies a numerical reference for reviewing your own generated draft. It is not a product benchmark. Two tools can produce different visual styles while preserving exactly the same data, and identical-looking outputs can still contain different mistakes.

Next inspect labels and conventions. Are the groups in the intended order? Does the axis state the unit? If you choose error bars for another dataset, specify whether they represent standard deviations, standard errors, or confidence intervals. These quantities answer different questions and should never share an ambiguous “error” label.

Only after these checks should you refine colors, typography, spacing, and legend position. Ask for readable labels and distinguishable markers. Keep the numerical mapping unchanged throughout those edits.

5. Write the caption before the graphical abstract

A restrained caption for this example is:

Synthetic responses for two illustrative groups, with six observations per group. Points show individual values; horizontal marks show arithmetic means of 10 and 14 arbitrary units. No inferential test is reported. These data do not describe a real experiment.

The difference between the means is four arbitrary units. Relative to A’s mean, B’s mean is 40% higher: (14 – 10) / 10 × 100. That descriptive calculation does not establish statistical significance, treatment efficacy, or a causal effect.

Writing this caption first helps establish a boundary for the visual summary. A graphical abstract should not turn “a higher mean in a synthetic example” into “a treatment improves outcomes.” Nor should it add a mechanism diagram merely because a mechanism would make the composition more interesting.

6. Turn the checked result into a visual summary

Use SciDraw AI’s graphical abstract maker to draft a composition from a short, verified description. The page accepts a written research summary and also offers a PDF-to-prompt route. For this exercise, text is sufficient.

Keep the brief compact and explicit:

Draft a three-panel graphical abstract for a synthetic teaching example. Panel 1: two independent illustrative groups, six observations each. Panel 2: measurement of response in arbitrary units. Panel 3: space for a separately verified dot plot, with the text “Mean A = 10; Mean B = 14”. Use arrows for workflow order only. Do not depict a biological mechanism, significance stars, or a proven treatment effect. Include “Synthetic example — not experimental evidence”.

Figure 3. An editorial storyboard for the graphical abstract. The result panel reserves space for the checked chart. This AI-generated editorial illustration is a planning aid, not a SciDraw AI output. The blank result panel is intentional: the verified chart is added during final assembly.

Review the generated text character by character. Check every number and arrow. If an illustrative layout contains an invented mini-chart, replace it with the verified chart during final assembly rather than treating the illustration as a numerical source. Preserve a separate copy of that chart so collaborators can inspect it independently.

For a real paper, work from the approved analysis and manuscript. Include only methods and findings actually supported there. Check the target journal’s current requirements for dimensions, file formats, and AI-assisted imagery before preparing the submission version.

7. Leave a reviewable handoff

Package the source CSV, final chart, graphical abstract, caption, and a short note describing any transformations. Record which values were checked and which parts of the composition were AI-assisted. Do not rely on a polished image alone as the record of the work.

A colleague should be able to answer three questions without reconstructing your session: what data went into the figure, what each visual mark represents, and which conclusions the figure supports. If those answers are clear, AI assistance has helped the communication process while leaving the scientific judgment with the researcher.