Image to Text

Overview

There are two ways to get text out of a picture, and they fail differently. An OCR engine recognizes characters: it is local, repeatable and reports a confidence per word, but it does not understand layout and it swaps look-alike characters. A vision model reads the way a person does: it follows tables, forms, angles and handwriting, but it gives no confidence and can quietly tidy, skip or invent text. Neither failure announces itself. In the runs behind this skill Tesseract returned a wrong date digit at confidence 83 and a wrong card digit at confidence 80. So the job has three parts: pick the route from what is in the image, read, then verify with a signal that does not depend on the route.

Instructions

1. Look first, then choose the route

Open the image with the agent's file-reading tool so the model sees the picture. Note the kind of content, the layout, the quality (blur, angle, lighting, text height) and the language. Ask what the text is for, which values must be exact (amounts, dates, codes, names), and whether the image may be sent to a hosted model at all: identity documents, medical and financial papers often may not.