Back to Liquid Agent DOCUMENTATION

Questions for Real Research Situations

Use a question that states your scientific purpose, what may run, and where to stop. The first seven scenarios below follow the real GSE174302 browser walkthrough; its screenshots show the actual questions, API answers and local outputs. The additional assay examples are adaptable prompts, not claims that those studies were executed in this documentation review.

1. New to the dataset: inspect, then plan

Situation: You have two circulating-RNA count matrices and want to know what can be learned before spending time on analysis.

I am studying circulating RNA in colorectal cancer. Inspect the attached GSE174302 cohort and its two count matrices. Explain the sample groups and measurement differences, then propose a staged plan: QC each matrix, compare CRC with healthy samples, compare the two sets of results, and write an exploratory report. Use appropriate skills. Do not run the analyses yet.

Inspect: Sources should show the intended cohort; Plan should name actionable steps and prerequisites. This request did not execute an analysis in the live run. Read the plan before clicking Run next step. Screenshots.

2. Continue a multi-step study without repeating work

Situation: The first QC is complete. You are ready to authorize the remaining calculations.

Proceed with QC of the second matrix and the CRC-versus-healthy differential analysis of each matrix. Then inspect the available comparison tasks and run matched-expression and differential-effect concordance. Keep the two measurement definitions separate, preserve model warnings, and publish an exploratory report. Reuse completed QC; do not rerun it.

Inspect: New work should correspond to remaining steps. In the live run, there were six scientific runs in total, not a new QC every time a question was asked. The same 73 samples in two measurement definitions do not constitute independent validation.

3. A figure looks unusual: select its actual data

Situation: Several PCA points lie far from the main cluster. In Results, drag a region around them and enter:

Inspect these PCA outliers in the source data. Are they technical or biological? Do not exclude samples.

Inspect: Does the popover say exact mapped marks or image only? The live selection mapped six samples. The answer inspected source values but could not determine the cause or justify exclusion. Selection and answer screenshots.

4. Compare evidence across two figures

Situation: You want to relate a PCA view to a heatmap block. Stage one crop from each by leaving the region field empty and clicking +, then ask:

Compare the PCA outlier region with this heatmap block. Inspect both mappings; explain their units and whether they identify the same samples. Summarize the selected data separately and write a short follow-up report. Do not merge selections or rerun differential analysis.

Inspect: Both images should appear in your message. In the real example, the second region mapped 165 cells; a local selected-data table and report were produced. The agent did not claim the two crops identified the same individuals. Two-image message and table.

5. Understand an unfamiliar figure from elsewhere

Situation: You have PNGs, but no underlying dataset. Upload them with +, then ask:

These two figures come from circulating-RNA QC. I am new to this: what is the difference between the PCA scatter plot and the feature heatmap? Does either prove that cancer can be diagnosed? Explain the colour scale and what information is still missing.

Inspect: A useful answer explains axes, units and limits without inventing source-data access. This was a real two-image API question in a separate chat. Uploading a PNG does not recreate a mapped Results selection. Upload and answer.

6. Teach a reporting preference without silent changes

Situation: You want a consistent report style for future work.

For future circulating-RNA reports, start with three plain-language findings and put comparison figures before dense tables. Explain that PCA is exploratory and that heatmap colours are relative z-scores in figure captions. Please remember these preferences, but show me proposed skill edits before applying anything.

Inspect: Read the reason and red/green diff. Accept only the changes you want. A one-off instruction need not create a permanent preference. In the real example, two drafts appeared; one was accepted and the other rejected. Actual review.

7. Review and correct a report using retained evidence

Situation: The report exists, but the order or measurement labels need attention.

Regenerate the full report using the accepted reporting guidance and the completed analyses. Include both comparison figures and retain fit warnings and study limitations. Do not repeat matrix calculations.

Then, when labels are ambiguous:

Verify which differential output belongs to each input definition. Check feature counts and finite estimates against the corresponding source provenance. Correct swapped table captions without recomputing.

Inspect: Read the tables as well as the answer. During this documentation run, report review caught swapped measurement captions; an explicit provenance-based correction was needed. A successful model turn is not a substitute for scientific review. Report review.

Additional tasks to adapt to your own data

These prompts describe intentions. Availability depends on actual inputs, installed tools, model weights and study design. Ask for inspection first; do not assume a suggested operation has already run.

Situation	Example request	Check before proceeding
Two separate datasets	“Attach these two sources. Keep their analyses separate and propose a defensible comparison; do not pool them automatically.”	Join keys, assay units, cohorts and batch confounding.
Missing labels	“Assess QC without group testing. Tell me which metadata would make a comparison valid.”	Do not infer outcomes from short sample codes.
Paired/longitudinal data	“Check patient and draw identifiers. Explain whether the available tool supports the repeated-measure design before proposing a contrast.”	Independent-sample testing may be inappropriate.
cfDNA raw tracks	“Inspect the coverage tracks, run the available numerical summary and visualization, and explain what cannot be inferred without aligned-read information.”	Raw coverage summaries are not a validated CNV caller.
Encoding	“Inspect format and local encoder availability; estimate the required preparation before running feature extraction. Do not download weights or launch encoding yet.”	Compatible inputs, installed weights, resources and time.
Existing embeddings	“Use the existing feature store to review outliers and a 2D projection. Do not encode the same files again.”	Features must exist and correspond to the intended samples.
Variant table	“Check the schema and filter support. Summarize VAF and recurrence, retaining germline/CHIP and depth limitations.”	Depth, annotation and matched-normal assumptions.
Methylation/CNV matrix	“Inspect the declared units, then propose QC and exploratory sample/region figures. Do not treat a signal matrix as a diagnosis.”	Normalization and biological interpretation depend on assay.
Digital PCR	“Quantify accepted partition counts with uncertainty, flag zero-positive and saturated wells, and publish the numerical table.”	This does not validate gating or assay detection limits.
Protein or EV table	“Summarize completeness and distributions before a group comparison; explain missingness and unit limitations.”	Missingness can confound comparisons.
Methods and literature	“Find evidence for a suitable liquid-biopsy method, distinguish papers from local findings, and cite only sources actually retrieved.”	Literature retrieval uses the network and does not validate this cohort.
Interrupted task	“Summarize what completed, what failed and what remains. Continue only unfinished authorized work from retained results.”	A cancelled job is not a completed result.

For schemas and implemented operations, use assay tables, cfDNA analysis, encoding and the capability matrix. The illustrated walkthrough provides a complete example.

Interface shortcuts

• Ask explains the selected result; Use for next step requests follow-up advice; Run next step executes a plan step.
• Guide steers a running turn; Stop requests cancellation. Read the resulting status before continuing.
• Default Web: liquid-agent; explicit terminal: liquid-agent cli; public homepage: liquid-agent wiki.
• In CLI, /image "/path/to/figure.png" What does this show? is an explicit image request. See CLI commands for key and skill-review management.