Your first task

The fastest way to learn Interpreter is to give it one boring job you already do by hand. This page walks through a real one: pulling fields out of a folder of PDFs into a spreadsheet.

Pick a workflow you already do

Pick something repetitive, file-based, and easy to verify. Good first tasks:

  • extracting fields from invoices, receipts, or statements
  • renaming and sorting a folder of downloads
  • reformatting rows in a spreadsheet
  • pulling a few values out of each PDF in a folder

Avoid anything that sends email, posts to a portal, or moves money on the first run.

Set the workspace

Point the workspace at the folder that contains both the source files and the destination file. For this walkthrough, set the workspace to ~/Documents/invoices-q1/ and put an empty invoices.xlsx inside it.

The agent can only touch files inside this folder. If it needs anything outside, it will ask.

Write the instruction

Be specific. Name the input folder, the output file, the exact columns, and a stopping condition.

Read every PDF in ~/Documents/invoices-q1/, extract vendor, invoice number, date, and total, then append rows to invoices.xlsx. Stop before sending any emails.

Notes on what makes this work:

  • It names a concrete input folder and output file.
  • It lists the fields explicitly, in order.
  • It ends with a hard stop ("Stop before sending any emails") so the agent does not keep going past the part you want to verify.

Watch the first run

Run it on one or two files first. Keep the overlay visible.

  • Interpreter will read the first PDF, open the spreadsheet, and start filling rows.
  • It will pause for approvals before writing the file the first time, and before anything irreversible.
  • If a column is wrong (date format, total includes tax, vendor truncated), stop the run and edit the instruction. Add an example row if it helps.

Repeat on the batch

Once one file looks right, tell Interpreter to continue with the rest of the folder. For larger batches, ask it to log any PDF it could not parse to a skipped.txt so you can review them at the end instead of stopping the run.

Next: Connect a model.

Give feedback