Make a request to a sighted AI to transfer a table with a photo into a working format. Return JSON {"prompt":"...","columns":"...","format_rules":"..."}: prompt - task, columns - which columns and in what order the AI should return, format_rules - in what form to output the result and how to deal with problem areas of the table.
The table in the photo is a grid that AI must not only read, but also save as a grid: rows, columns, cells in their places. There are two enemies here. The first is skew and shooting angle: the rows “float”, and AI can shift the value to the adjacent column. The second is merged and empty cells: the model strives to “think through” them. Pro trick: hard set the output format and column headers in the desired order - for example, “return as CSV with columns Date, Customer, Product, Quantity, Amount.” CSV is a simple text type of table, where the values in the row are separated by commas (or you can simply ask “return as a table” - the main thing is that the structure is specified explicitly). The finished format is inserted into the table without edits. The second technique: write down the rules for the edges - leave the empty cell empty, duplicate or mark the combined one, keep the number of lines. Hidden control lever: Ask the AI to tell you how many lines it has counted, and mark the lines it read with uncertainty. Coincided with the original - fair transfer. A typical beginner’s mistake: not specifying either the format or the columns - the AI produces the text “in a sheet”, and then you still have to parse the data by hand.
Unlock access to submit solutions for instant AI review.