Project: release an accessible transit disruption alert

Last updated: 5 Oct 202619 min read

project

A fictional transit agency must publish a Line K disruption alert. Its release packet contains a diversion map, a 74-second spoken update, a stop schedule, a claim notice, and two localized versions. Build a prompt contract and review path for each artifact. Keep approved service facts as typed records: Alder and Birch are closed; the shuttle boards at Cedar; the claim applies only to riders meeting the payment and disruption windows. No assistant output becomes public until its claims and final rendering are checked.

Prepare the artifacts

Classify the route icon as decorative in the heading context, then create a short map alternative and nearby route detail from the approved map. Draft timed captions from the recording, preserving the overlap at 00:39 and the service chime at 00:41.500. Produce a transcript that also explains the on-screen diversion diagram. Send stop and departure rows to a renderer that creates real table headers rather than asking the model for aligned spaces. Rewrite the claim notice in plain language without widening eligibility or promising immediate payment.

Check the published task

Test a reader trying to find the shuttle stop, hear or read the alert, inspect the next departure, and file a valid claim. Include one control with an unreadable map label and another with an unknown spoken phrase; the model must mark uncertainty instead of inventing facts. Compare the localized notices with the same eligibility record. Inspect the rendered page with keyboard and assistive technology as well as structural checks. Record the image version, media hash, renderer build, reviewer, failures, and release decision. A false boarding stop blocks release even when every other fixture passes.

Map: Alder/Birch closed; board shuttle at Cedar.
Audio: overlap 00:39; chime 00:41.500; unresolved words stay marked.
Schedule row: Cedar | 08:47 | shuttle; renderer owns headers.
Claim: paid before 08:30; affected 08:45–10:15; file within 47 days.
Gate: verified facts + rendered task check; false stop -> block.

Performance and operating cost

Let A be artifact fixtures, V prompt variants, and R review modes. The broad evaluation is O(AVR), plus O(D) playback time for total media duration D and a structural walk linear in document nodes. Reuse typed facts and approved versions across drafts; this prevents a map revision from leaving stale prose behind. Keep separate defect counts for unsupported visual facts, caption omissions, structure, misleading claim conditions, and locale drift. Human playback and task review are slower than a schema check, but a schema check cannot hear an overlapped warning or judge whether the page makes the shuttle stop findable.

Common Mistakes

Do not publish a model's map description without comparing it to the approved asset.
Do not confuse a transcript with synchronized captions.
Do not infer accessibility from a passing prompt test when the published template has not been inspected.

Connected lessons

prompt engineering

accessible output