Run one Claude prompt across every row of a Google Sheet

Build me an internal app my ops team opens whenever we have a spreadsheet of rows that all need the same Claude treatment. The recurring jobs are things like classifying support conversations into themes, writing product descriptions from a list of attributes, and summarizing free-text survey answers. Today that means pasting rows into a chat window one at a time, so the app has to cover the whole ritual: set up the job, submit it, watch it, review what came back, and write the good answers into the sheet.

The main screen is New Run. I paste or pick a Google Sheets spreadsheet, choose the tab and the range, and the handler calls Get Values to read it. Show me the header row and the first several rows in a small preview table, then let me choose which single column feeds the prompt. Below that, a prompt template box where I write one instruction and reference the chosen column with a placeholder, for example {{column}} or the header name. Substituting each row's cell into that template is what produces one request per row.

Next to the template, a model picker populated by List Models, so the list stays current rather than being hardcoded. Beside it, an estimate panel: take three sample rows, build the exact request each would produce, and call Count Message Tokens on them. Show the estimated input tokens per sample row, the average, and that average multiplied by the total row count as a rough job size. Label it clearly as an estimate of input only. We have no usage or cost reporting available, so do not build a spend dashboard, do not show dollar figures, and do not promise invoice-level tracking. The token estimate is the honest ceiling, and the one cost fact worth stating is that batch requests are billed at half the standard rate.

Submitting sends every row as one job through Create Message Batch, with each request carrying a custom_id built from the sheet row number so results can be matched back later. A batch accepts up to 10,000 requests, so if the selected range has more rows than that, split it into successive batches and group them under one run in the UI. Warn me before submitting anything over a few thousand rows. After submitting, take me to the status board.

The status board lists every job. Use List Message Batches for the overall list and Retrieve Message Batch for the live counts on each one, showing processing status plus how many requests have succeeded, errored, been canceled, or expired, as a progress bar. These batches are asynchronous and can take up to 24 hours, though most finish in under an hour, so the board must refresh when I open the page or click a refresh button. Do not imply anything runs on a schedule and do not build background polling. Each in-progress job gets a cancel button wired to Cancel Message Batch, with a confirm step. Ended jobs can offer an optional cleanup action using Delete Message Batch.

When a job has ended I open its review screen. Fetch the outputs with Get Message Batch Results, which returns JSONL, and join each result back to its original row using the custom_id. Render a table with the original cell text in one column and Claude's answer beside it, one row per input row, with a checkbox on each. Rows that came back as errors show the failure inline and stay unchecked rather than being hidden, so nobody thinks a row silently succeeded. Everything else starts checked. Let me untick anything bad and edit an answer in place before approving.

Approving writes only the checked rows back to the sheet. Give me two destinations: write into a new column beside the original rows using Update Values, aligning each answer to its own row number so nothing shifts, or add the results as new rows on a separate tab using Append Values. Ask which one before writing, show how many rows are about to be written, and report how many landed.

Persist a history of past runs in app storage, each saved with the spreadsheet and tab, the column that was used, the full prompt template text, the model, the batch ids, the row count, the submitted and finished timestamps, how many results were approved, and who submitted it. Everyone in the workspace sees all runs. Opening a past run shows its prompt and results, and a rerun button refills the New Run form with the same template, column, and model so the same job next month takes one confirmation instead of a rebuild.