PDF → spreadsheet

Reading PDFs and turning them into structured data is one of the highest-volume things LMCP users do. This skill codifies the end-to-end flow: read the PDF(s) on the user's computer (macOS or Windows), extract the rows, and build an Excel sheet. The tools execute on the device; the extracted text goes to the user's AI provider as part of the conversation.

The flow

1. Read the PDF(s) — pdf_read(path=...), one call per file. (Find the files first with file_search, or finder_search on macOS, if the user only describes them, e.g. "my March invoices".)
2. Extract the structured rows from the text — identify the columns the user wants (e.g. for statements: Date, Description, Amount; for invoices: Vendor, Invoice #, Date, Total). Normalize dates and numbers. If several PDFs, combine into one consistent table.
3. Write the spreadsheet — prefer one call: excel_create(path="<out>.xlsx", headers=["Date","Description","Amount"], rows=[[...],[...]], confirm=true). headers is required; rows is the full data. For edits to an existing sheet use excel_write_cell(path, cell="B2", value, confirm=true); to read one back use excel_read(path).

Good practice

• Confirm the columns with the user before building the sheet ("I'll pull Date, Description, Amount — want a Category column too?").
• Build the whole table in memory, then write it in a single excel_create (one row per record) — don't write cell-by-cell unless editing an existing file.
• Watch totals: if the PDF has a stated total, add a check row or verify your sum matches.
• Keep money as numbers (not "$1,234.50" strings) so the sheet can compute.
• Write the output next to the source or to a path the user names; excel_create/excel_write_cell are preview/confirm-gated (pass confirm: true once the user approves the columns).
• Multi-PDF: keep a Source column so each row traces back to its file.