How to Prompt

You do not need to write a formal protocol before starting. A strong request makes the scientific goal and evidence boundary clear, then lets Biomni plan the implementation.

Start with the outcome

Say what you need to learn, decide, or deliver.

Too broad
Actionable

Analyze my data.

Identify genes that differ between treatment and control in @counts.csv,
then create a ranked result table and a publication-ready volcano plot.

Include the experimental context

Add details that change the interpretation or method:

organism, reference build, assay, and tissue or cell type;
cohort definitions, controls, replicates, and batch structure;
units, column meanings, and missing-value conventions;
primary endpoint or research question; and
whether the work is exploratory, confirmatory, or intended for publication.

These are paired tumor and adjacent-normal RNA-seq samples from 12 patients.
Patient ID is the blocking variable in @metadata.csv. Use GRCh38 annotations and
report both effect size and adjusted p-value.

Attach or mention the evidence

Use + Add for files, folders, cloud imports, Skills, resources, and connectors. Type @ to search files and scientific resources, or + to search Skills. Do not rely on a filename alone when its structure is ambiguous. Explain which sheet, columns, groups, or identifiers matter.

State constraints without over-specifying

Specify a method when it is scientifically required. Otherwise, state the constraint and ask Biomni to choose and justify the method.

Use a method appropriate for paired count data. Before running it, explain the
design formula, filtering rule, and multiple-testing correction you chose.

Useful constraints include:

required or prohibited methods;
quality thresholds and inclusion criteria;
resource or runtime limits;
databases or date ranges to search;
output format, dimensions, and naming; and
citation or reproducibility requirements.

Ask for inspectable outputs

Request the artifacts needed to audit or continue the work:

Return the full result table as CSV, the analysis notebook, SVG and PNG versions
of each figure, and a short README describing inputs, software, parameters, and
output files.

For literature or database research, ask for linked sources next to the claims they support. For data analysis, ask Biomni to separate observed results from interpretation.

Use follow-up turns

Keep related refinement in the same task so Biomni can use the conversation and generated files.

The plot is correct, but label only the ten genes with the smallest adjusted
p-values and increase the axis-label size. Do not rerun the statistical model.

If the goal changes substantially, start a new task. This keeps the evidence, assumptions, and results for each analysis easier to review.

When you are unsure

Ask Biomni to identify missing information before it acts:

I want to test changes across three timepoints in this RNA-seq experiment. Read
@metadata.csv and tell me what design details you still need before choosing a
model. Do not run the analysis yet.

For high-impact conclusions, finish with an explicit verification request and run Scientific Review.

Get Started

Work in Biomni

Data and Teams

Account

Support

Start with the outcome

Include the experimental context

Attach or mention the evidence

State constraints without over-specifying

Ask for inspectable outputs

Use follow-up turns

When you are unsure

​Start with the outcome

​Include the experimental context

​Attach or mention the evidence

​State constraints without over-specifying

​Ask for inspectable outputs

​Use follow-up turns

​When you are unsure

Start with the outcome

Include the experimental context

Attach or mention the evidence

State constraints without over-specifying

Ask for inspectable outputs

Use follow-up turns

When you are unsure