A library agent must read a scanned community-event flyer before answering patron questions about the event. Which tool pattern is appropriate?

A library agent asked about a community event on a scanned flyer should use a content-understanding tool on the image first, then answer from the extracted content. Scanned flyers are visual artifacts; without OCR/document understanding, the language model has no reliable access to dates, venues, or contact lines printed on the page. Extract-then-answer keeps patron guidance tied to what the flyer actually says rather than using an image-generation tool on the scan, which generates pixels rather than extracting readable dates, venues, or contact lines to answer from. Using an image-generation tool on the scanned flyer and answering from the extracted content fails because generation produces another image without extracting machine-readable text, so there is no reliable extracted content and patrons could be sent to the wrong time or place if the recreation distorts details. Running an audio-transcription tool on the scanned flyer fails because transcription expects spoken audio, not a visual page, so it cannot extract the printed dates, venues, and contact lines, leaving no extracted content to answer from. Running a text-translation tool on the scanned flyer fails because translation converts already-extracted text between languages rather than reading text from a scanned image, so it cannot extract this flyer's actual dates, venues, or contact lines for the answer. On the exam, content-understanding tools bridge scanned civic materials into agent answers; distractors use image generation without extracting content, use audio transcription on an image with no audio, or use text translation without reading this scanned notice. Caveat: extraction quality varies with scan quality—low-contrast flyers need confidence checks or staff review before publishing event answers to patrons. Operational check: submit a scanned event flyer, run content understanding, confirm key fields (date, time, location) are extracted, and verify the patron answer cites those fields rather than model invention.