Sheet Agent

Problem and Use Cases

When working with CSV or .xlsx spreadsheets, the hard part is often not opening the file but answering business questions quickly: which orders have not been followed up for more than X days, which amounts exceed Y, which customers are VIP, and how to aggregate quantities by sales representative. Manual filtering, pivoting, and cleanup can interrupt the workflow, while direct model-driven edits may introduce mistaken writes, missing backups, or wrong assumptions about column meaning. sheet-agent targets these spreadsheet tasks by putting natural-language querying, anomaly checking, and change suggestions into a controlled workflow.

How It Works

The skill separates spreadsheet reading, business queries, anomaly detection, and pre-write confirmation. It supports CSV and .xlsx, recognizes column names and data types, and attempts to infer the business context, such as orders, inventory, leads, or daily reports. During querying, natural language can be turned into structured results, such as “orders with amount greater than 10000,” “quantity totals by sales representative,” or “all VIP customers.” Anomaly detection covers missing values, negative numbers, duplicate IDs, inconsistent date formats, and values outside reasonable ranges. For daily or weekly summaries, it detects the time range and outputs overall counts, averages, distributions, top items, anomaly records, and trend descriptions.

For write operations, the key mechanism is preview first: when a user requests a change, the agent shows a preview without writing yet; it executes only after confirmation and does nothing if the user cancels. It is read-only by default, automatically backs up the original file to the backup/ directory before writing, and asks when column meaning is unclear instead of guessing.

Boundaries and Caveats

This skill is best suited for single-file spreadsheet querying, inspection, and summary generation. The docs recommend a maximum of 100,000 rows per operation and support CSV and .xlsx; version 1 does not handle cross-spreadsheet joins. If a task requires multi-table joins, complex modeling, or high-risk bulk writes, treat it as an assistive analysis tool and keep manual review and backup controls.