Alternative text describes an image's purpose in its page context, not every visible pixel. A prompt must receive the image, its nearby copy, the destination or action if the image is a control, and a declared role: decorative, informative, functional, or complex. The generator should return a proposed short alternative plus a reason and an uncertainty marker. A decorative image can receive an empty alternative in the rendered page. A complex chart needs its important relationships in nearby text or a data equivalent; a crowded one-sentence alternative cannot carry the entire analysis. No image evidence means no visual claim. A human editor chooses the role and checks the rendered result.
Image alternatives: prompt for purpose, then check the image
Operational case
A fictional transit bulletin shows a small blue route icon beside the heading 'Line K diversion'. Because the heading already carries the route identity and the icon adds no information, the editor marks it decorative. A second graphic is a route map: Stops Alder and Birch are closed, and a shuttle serves Cedar. Its adjacent bulletin text omits Birch. The assistant drafts a short map alternative and a separate textual route sequence; the reviewer compares both against the approved map and corrects the missing stop before release. If the icon becomes a link to the route status page, its alternative must name that destination instead of the color or shape.
INPUT: asset=MAP-47; context=Line K diversion; role=complex map.
Verified features: Alder and Birch closed; shuttle at Cedar.
OUTPUT: short_alt='Line K diversion map'; detail='Alder and Birch closed; shuttle boards at Cedar.'
REVIEW: compare detail with approved map; publish both in page context.Performance and operating cost
For I assets and an average of F verified features per asset, feature checking is O(IF) review work; generating a short sentence does not remove that work. Keep the approved feature list beside the asset version so a revised map invalidates its old description. A vision model may reduce drafting time, yet it can miss tiny labels or infer a nonexistent connection. If the map cannot be read with enough certainty, request the underlying route data. The site editor must verify the actual HTML alternative and the adjacent detail, not only the model's proposed string.
Common Mistakes
- Do not describe a decorative icon as if it adds new information.
- Do not use a link image's appearance in place of its destination or action.
- Do not compress a complex map into a short alternative without a usable detail equivalent.
Connected lessons
- Prompt engineering applications
- Prompt Engineering
- Image prompts: verify claims against named regions and crops
- Image prompts: specify the visual contract and inspect the pixels
- HTML Accessibility
- HTML figure and figcaption: keep an asset with its explanation
- Caption and transcript prompts: keep timing, speakers, and sound evidence
- Semantic output prompts: verify structure after generation
- Plain-language prompts: simplify without losing conditions
- Accessible output evaluations: measure failures by artifact and user task
- Project: release an accessible transit disruption alert
- Accessible prompt output decisions
Continue with: Editable report prompts: inspect structure and rendered pages.
Continue with: Chart prompts: review rendered and nonvisual outputs.
