excel

name: excel description: Create validated Excel .xlsx workbook artifacts with XlsxWriter, formula-injection protection, and Rockie artifact emission.

excel

Use this skill when the user asks for an Excel workbook or an .xlsx artifact, including creating, modifying, formatting, analyzing, validating, or exporting workbook content. For read-only analysis or validation requests, inspect the workbook and report findings without writing a new file unless the user asks for one.

In tenant runtimes, new workbook generation is productized through skills/excel/runtime/renderer.py. Generate the workbook locally in the tenant workspace, verify it, then publish it with the emit_artifact tool using:

• kind: "spreadsheet"
• content_encoding: "base64"
• filename: "<safe-name>.xlsx"
• mime_type: "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet"
• destinations: ["ui", "chat"]

Core rules

• Produce a real .xlsx workbook. Do not rename CSV, TSV, HTML, JSON, or plain text files to .xlsx.
• Use runtime/sanitize.py at the cell-write boundary. Strings starting with =, +, -, or @ are text unless the cell is explicitly typed as a formula.
• Preserve source data. Keep raw inputs available when appropriate, such as a Source Data or Raw Data sheet, and avoid silently replacing user-provided values with summaries.
• Use workbook features that fit the task: multiple sheets, formulas, number formats, styles, column widths, freeze panes, filters, charts, data validation, named ranges, and notes/comments when they materially improve usability.
• Keep formulas as formulas unless the user explicitly asks for static values.
• When writing a workbook, reopen and verify the finished file before reporting it.
• When writing a workbook, report the emitted artifact id or final path.
• Cap new workbook generation at 20 sheets. If the user asks for more, create the most useful 20 or fail clearly before writing, depending on the task.

Verification

After writing the .xlsx, reopen it with an Excel-capable parser or by inspecting the ZIP/XML package when no dependency is available. Verify:

• Expected workbook and worksheet names.
• Sheet dimensions and important ranges.
• Target cells changed as requested.
• Formulas are present as formula XML or formulas, not only cached values.
• Styles, number formats, widths, freeze panes, filters, charts, and validation rules that matter to the task.
• Sample values from source data and calculated or summarized outputs.

If verification fails, fix the workbook and verify again before responding.

Runtime renderer

The renderer accepts a JSON request and writes a workbook:

python "${OPENCLAW_SKILLS_DIR}/excel/runtime/renderer.py" \
  --input request.json \
  --output workbook.xlsx

Request shape:

{
  "sheets": [
    {
      "name": "Sources",
      "headers": ["Title", "Type", "Score"],
      "rows": [["Paper A", "paper", 0.92]],
      "freeze_panes": "A2",
      "autofilter": true,
      "charts": [
        {
          "type": "column",
          "title": "Scores",
          "categories": "$A$2:$A$10",
          "values": "$C$2:$C$10",
          "position": "E2"
        }
      ]
    }
  ]
}

Cells can be objects when type matters:

{"type": "formula", "value": "SUM(C2:C10)"}

String cells that look like formulas are sanitized automatically unless they use the explicit formula object form.