What this prompt does
This prompt builds a spreadsheet analysis tool powered by Gemini, aimed at non-technical users who want to ask questions of their data in plain English. It runs through six steps: ingesting and profiling the file, building a natural-language query interface, cleaning data with Gemini assistance, generating automated insights, recommending visualizations, and exporting a report. Returning the generated code alongside each answer matters so users can verify results rather than trust a black box.
The variables shape the tool. [data_type] and [model_version] set the context, [file_formats] and [parsing_library] plus [preview_rows] drive ingestion, and [sdk_language] is the code returned with each answer. [date_format], [fill_strategy], [dedup_columns], and [outlier_method] govern cleaning, [max_rows] and [insight_count] shape insights, and [chart_library], [chart_count], and [export_formats] drive visualization and reporting. The cleaning variables especially deserve attention, since defaults like median fill are not always right for a given column.
When to use it
- You are building an internal tool for non-technical users to query spreadsheets in plain English.
- You want automatic profiling of column types, data quality issues, and a dataset summary.
- You need natural-language questions converted into aggregations, filters, and pivots.
- You want Gemini to suggest data-cleaning operations with previews before applying.
- You want automated insights and recommended chart types from the data.
- You need an exportable report combining quality overview, insights, and visualizations.
Example output
Expect an ingestion pipeline that parses [file_formats] via [parsing_library] and profiles the first [preview_rows] rows for column types, header location, and quality issues. The query interface returns both results and the [sdk_language] code that produced them; the cleaning step proposes operations for dates, missing values, and [dedup_columns] duplicates with row previews; the insights step surfaces [insight_count] trends; and the visualization step recommends chart configs for [chart_library]. A report exported in [export_formats] finishes it, optionally on a recurring schedule that highlights changes since the last run. It blends pipeline code and structured output, with the generated analysis code surfaced at every step so results stay auditable.
Pro tips
- Always surface the
[sdk_language]code with each answer so users can verify and modify the analysis instead of trusting opaque results. - Send only
[preview_rows]for profiling and sample large files beyond[max_rows], since sending full big datasets is costly and may exceed limits. - Set
[date_format],[fill_strategy], and[outlier_method]to match your domain; defaults like median fill are not always appropriate. - Preview affected rows before applying any cleaning operation, especially deduplication on
[dedup_columns], to avoid silent data loss. - Treat generated insights as starting points; harden the data-quality logic yourself before relying on the numbers.
- Pick a
[chart_library]your stack already supports so the generated chart configs run without extra integration or rework.
