Sheet Agent

Spreadsheet operations agent for CSV and Excel files: inspect data, detect anomalies, answer business questions, generate reports, and preview edits safely.

Security Scan

The skill's code, instructions, and requested access are internally consistent with a spreadsheet-analysis agent: it reads CSV/XLSX files, detects anomalies, previews changes, and only writes after confirmation (backing up to a local backup/ directory).

Purpose & Capability

Name/description match the provided code and tests: the package reads CSV/.xlsx, infers table types, detects anomalies, answers NL queries, generates reports, and previews/writes changes. No unrelated credentials, binaries, or services are requested.

Instruction Scope

SKILL.md and README limit operations to reading local spreadsheet files and producing previews/reports. The documented safety rules (read-only by default, explicit confirmation for writes, backup before write) are implemented in code and tests. Instructions do not ask for unrelated system files, environment variables, or external endpoints.

Install Mechanism

No install spec provided (instruction-only skill) and the code is included directly. This is low-risk compared with remote download installs. The script checks for Python deps (pandas, openpyxl) and exits if missing; dependencies are standard and expected.

Credentials

No environment variables, credentials, or config paths are requested. The skill operates on user-provided file paths only. There are no API keys, tokens, or unrelated env requirements.

Persistence & Privilege

always:false (normal). The skill will create a backup/ directory alongside the skill code and copy backed-up spreadsheet files there — this is a documented behavior and logically related to its purpose, but it does mean the skill will write files under its own skill directory when performing confirmed writes/backups. It does not modify other skills or global config.

Assessment

This skill appears coherent and implements the spreadsheet-reading, anomaly detection, reporting, and safe-change-preview features it claims. Before installing: (1) confirm you trust the unknown author (source/homepage are absent); (2) be aware the skill will read whichever local spreadsheet files you point it at and, upon confirmed edits, will create backups under the skill's backup/ directory (check where that folder will live in your environment); (3) the script requires pandas and openpyxl — install them in an isolated Python environment before running; (4) if you need stronger assurance, review the full scripts (they are included) or run the bundled test suite in an isolated environment to validate behavior. No network calls, credential access, or obfuscated code were found in the provided files.

One-Line Positioning

Transform natural language into spreadsheet comprehension, anomaly detection, business insights, and safe rewrite suggestions.

---

When to Use

Activate this skill when the user mentions:

• Having AI read or inspect a spreadsheet for problems
• Querying data from CSV/Excel files
• Generating daily reports, weekly reports, or operations summaries
• Modifying a specific record in a spreadsheet (preview first, confirm second)
• Checking for empty values, duplicates, or anomalies in a spreadsheet

---

Usage

/sheet "spreadsheet path" "natural language instruction"

Examples:

/sheet "~/Desktop/orders.csv" "Which orders haven't been followed up for more than 3 days?"
/sheet "~/Documents/inventory.xlsx" "Check if there are any negative inventory quantities"
/sheet "~/Desktop/orders.csv" "Generate last week's weekly report"
/sheet "~/Desktop/customers.csv" "Change the customer tier in row 8 to VIP"

---

Core Features

1. Spreadsheet Reading & Comprehension

• Supports CSV and Excel (.xlsx) files
• Auto-recognizes column names and data types
• Infers business type (orders/inventory/leads/daily report)

2. Business Queries (Natural Language to Structured Results)

• "Orders not followed up for more than X days"
• "Orders with amount greater than Y"
• "Find all VIP customers"
• "Sum quantities by sales representative"

3. Anomaly Detection

• Empty values (missing data)
• Negative numbers (inventory/quantity columns)
• Duplicate IDs
• Date anomalies (future dates, inconsistent formats)
• Value anomalies (beyond reasonable ranges)

4. Daily / Weekly / Operations Summaries

• Auto-detects time range
• Generates structured summary:
  • Overall view (total count, sum, average)
  • Distribution stats (category percentages, top items)
  • Anomaly record list
  • Trend descriptions

5. Change Preview (Key Safety Mechanism)

All changes show a preview first — execute only after user confirmation.

User says "Change the amount in row 5 to 15000": -> The agent shows a preview and does NOT write yet -> If the user replies "confirm", execute the write -> If the user replies "cancel", do nothing

---

Output Examples

Query Results

📊 Query Results: 12 leads have not been followed up for more than 3 days

| # | Name | Phone | Entry Date | Days Since |
|---|------|-------|------------|------------|
| 3 | Alex Wang | 138... | 2026-03-25 | 6 days |
| 7 | Leo Zhao | 139... | 2026-03-27 | 4 days |

Anomaly Detection

⚠️ Found 4 anomalies:

1. Row 15 [Bluetooth Earbuds]: Inventory = -5 (negative)
2. Row 28 [Phone Case]: Unit price is empty
3. Row 42 [Data Cable]: Inventory = 9999 (suspiciously large)
4. Row 55: Product name is duplicated

Weekly Report

📈 Operations Summary — Week 4, March 2026

[Overall View]
- Total orders: 156
- Total sales: ¥128,500
- Average order value: ¥823

[Top 5 Selling Items]
1. Bluetooth Earbuds - 45 units
2. Phone Case - 32 units
...

[Anomalous Orders]
- Refunds: 2 orders

Change Preview

📝 Proposed Change Preview:

File: customers.csv
Row: 8
Field: Customer Tier
Old value: Regular
New value: VIP

⚠️ Confirm execution? Reply "confirm" or "cancel"

---

Safety Principles

1. Read-only by default: Any operation does not modify files by default
2. Write only on confirmation: Write operations require explicit user confirmation
3. Backup before write: Automatically back up original file to backup/ directory
4. Ask when uncertain: If column meaning is unclear, proactively ask the user instead of assuming

---

Limitations

• Recommended maximum of 100,000 rows per operation
• Supports CSV and .xlsx formats
• Version 1 does not handle cross-spreadsheet joins