Spreadsheets and data

Interpreter cleans, joins, and categorizes data across spreadsheets and PDFs in your workspace. Use it when data needs to move between formats, get normalized, or be classified against a policy you can name.

Cleaning a CSV

Ask for specific, named operations. Vague requests produce vague results.

  • Deduplicate rows by a column you name (e.g. email)
  • Normalize names to "First Last" capitalization
  • Split a single address column into street, city, state, zip
  • Flag missing values in required columns
  • Coerce date strings to ISO format

Always tell Interpreter where the output goes: overwrite the file, write to a new sheet, or write to a new file with a suffix.

Joining a PDF into a sheet

When data lives in a PDF report and needs to land in existing rows, name the exact fields and destination columns.

  • Source: q1-vendor-report.pdf
  • Match key: vendor_id in column A of the spreadsheet
  • Pull fields: total_spend, contract_end_date, account_owner
  • Destination columns: E, F, G

If a row in the sheet has no matching record in the PDF, tell Interpreter to leave the cells blank and add the row number to a missing tab.

Categorizing against a policy

Interpreter is strong at classifying rows when the policy lives in the workspace as a readable document.

Worked example

Categorize every row in corp-card-march.csv against the policy in expense-policy.md. Add a category and flag column. Flag anything over the per-meal limit or missing a receipt. Stop before emailing finance the summary.

Interpreter reads the policy, walks each row, assigns a category, and flags exceptions. It pauses before sending the summary email so you can audit the flags first.

Working over big sheets

Do not turn Interpreter loose on 10,000 rows. Run a small batch, review the result, then expand.

  • Ask for the first 5-10 rows.
  • Spot-check the category and flag columns.
  • If the logic looks right, tell Interpreter to continue against the rest.
  • If it is wrong, correct the instruction and rerun the small batch.

This loop catches misread fields and bad policy interpretations before they touch the whole file.

Require approvals at the right points

  • Before overwriting a source spreadsheet
  • Before emailing a summary to finance, ops, or a client
  • Before exporting flagged rows to a downstream system
  • Before deleting rows you marked as duplicates

Stop before any step that leaves your machine. Cross-link: /docs/desktop/approvals.

Give feedback