Feeding Excel workbooks to an LLM

MakeItMarkdown · July 2026

Spreadsheets look like tables but store something stranger: typed cells whose displayed value and stored value differ. That gap — dates stored as day-counts, formulas stored as expressions, layout stored as merges — is where models get systematically misled.

What breaks if you hand it over raw

Serial dates: the cell shows 2025-09-22, the file stores 45922. Raw extractions hand the model five-digit numbers it happily treats as amounts.
Formulas: depending on the path, the model sees =SUM(B2:B13) instead of the number — or a number with no hint it's computed.
Merged headers unmerge into one value and blanks, shifting columns.
Multi-sheet workbooks flatten into one undifferentiated stream, or lose all but the first sheet.

The element mapping

In the workbook	In the Markdown
Each sheet	Its own `##` section, in workbook order
Data range	GFM table with type-annotated headers — `units (int)`, `month (date)`
Date cells	ISO dates (`2025-09-22`), never serials
Formula cells	The cached computed value, with a note that values are cached
Rows beyond 50	Truncated with an explicit "first 50 of N" sentence
Ragged/merged regions	Padded to rectangular + a fidelity warning
Empty sheets	Named and flagged, not silently skipped
Charts, pivot caches, styling	Not extracted — values only (the report says what was detected)

Why the type annotations matter more than anything else here: Tables LLMs can actually read.

Before → after

In the file

A2: 45922            (formatted as 2025-09-22)
B2: =SUM(B3:B14)     (cached value 79)
C1: "Revenue
     (EUR)"           (merged across C1:D1)

In the Markdown

| month (date) | total (int) | revenue_eur (float) |
| --- | --- | --- |
| 2025-09-22 | 79 | 96432.10 |
Values are cached formula results from the last save.

Honest limits

Formulas are not recomputed — you get the value cached at last save. Stale saves mean stale numbers (the note in the output reminds the reader).
Charts don't survive: a chart is a drawing bound to data; the data survives, the picture doesn't.
Fifty rows per table is a deliberate paste-budget default — for full-table workloads, attach the .md and the original together, or ask the model for code to run over the real file.

FAQ

Legacy .xls? Yes — both the modern ZIP format and the legacy binary are read (each is verified by its magic bytes first, so mislabeled files fail loudly, not weirdly).

Several sheets, one question? Every sheet is a titled section, so you can tell the model "use the Summary section".

Financial data privacy? Local conversion, nothing uploaded — verifiably.

Convert a workbook and inspect the typed headers and the cached-value note.