Prompt Chain: Full-Product Accessibility Audit Across Every Screen

What This Builds

A three-step prompt chain that takes a full set of product screens, typically 10-25, and produces a single consolidated accessibility audit ranked by severity, instead of a one-off WCAG check on a single screen. The single-screen checklist from Level 1 tells you about one screen. This chain tells you about your whole product at once, with issues grouped so you can fix the pattern instead of chasing individual instances.

Prerequisites

• Plus subscription at $20/month, needed for image analysis across a full screen set in one extended session
• Screenshots or exported images of every screen you want audited, ideally at actual resolution rather than heavily cropped
• Familiarity with WCAG basics (contrast ratios, focus states, alt text) so you can sanity-check the AI's findings
• Total ongoing cost: $20/month a month, no separate cost per audit beyond your existing subscription

The Concept

A single-screen check is like proofreading one page of a document. This chain is closer to a full manuscript review: it reads every page, keeps a running list of the same mistake showing up in different places, and hands you one report grouped by issue instead of twenty separate page-by-page notes. The chain runs in three passes: screen-by-screen findings, cross-screen pattern grouping, then a prioritized fix list.

Build It Step by Step

Part 1: Prepare Your Screen Set

1. Export or screenshot every screen you want audited, at a size where text and UI elements are clearly legible, not a thumbnail.
2. Organize them in a logical order (typically the flow order a user would encounter them) and note a short label for each (e.g., "01-Login," "02-Dashboard," "03-Settings").
3. Keep the set to a manageable number per session, roughly 10-15 screens works reliably in one chain; split a larger product into two audit sessions if needed.

Part 2: Run the Chain

Prompt 1, Per-Screen Findings (run once per screen, or in small batches of 3-4 images):

Review this screen for accessibility issues against WCAG 2.2 AA. Check specifically: color contrast for text and interactive elements, whether interactive elements appear large enough to tap or click reliably, whether form fields appear to have visible labels, and whether any information seems to be conveyed by color alone. For each issue found, note: what it is, where on the screen, and estimated severity (Critical, Serious, Moderate, Minor). Label your screen "[screen label from Part 1]" at the top of the response.

Repeat for each screen or small batch, keeping every response in the same conversation so later prompts can reference all of them.

Prompt 2, Cross-Screen Pattern Grouping:

You've now reviewed all the screens I've shared in this conversation. Group the findings across all screens into recurring patterns rather than listing every individual instance. For example, if low contrast on secondary buttons showed up on 4 screens, report that as one pattern affecting 4 screens rather than 4 separate findings. List each pattern with: description, which screens it affects, and severity.

Prompt 3, Prioritized Fix List:

Using the patterns you just identified, produce a prioritized list ranking the top 5-8 issues to fix first. Prioritize by a combination of severity and how many screens each pattern affects. For each, suggest a specific fix direction (not final copy or exact pixel values, just the direction, e.g. "increase contrast ratio on secondary button text" or "add visible focus indicators to all interactive elements"). Flag any item that needs a developer to confirm feasibility versus something the designer can adjust directly in the design file.

Part 3: Verify Before Acting

Spot-check at least 3-4 of the flagged issues yourself against the actual screen and, where the finding is about contrast, run the exact color values through a contrast checker tool rather than trusting the AI's visual estimate. AI vision analysis of contrast from a screenshot is a strong starting signal, not a substitute for a real contrast ratio calculation.

Real Example: Auditing a 14-Screen Onboarding and Core Flow

Setup: A UX designer at a healthcare scheduling app exported 14 screens covering the full new-user onboarding flow and the core appointment-booking flow.

Input: 14 screen screenshots at full resolution, run through the chain in three batches of 4-5 screens for Prompt 1, then one consolidated Prompt 2 and Prompt 3 pass.

Output: The pattern-grouping step surfaced that low-contrast helper text appeared on 6 of the 14 screens, all using the same secondary text style, meaning one design-system token fix would resolve most of the flagged instances at once instead of 6 separate screen edits. The prioritized list put that fix first, ahead of several one-off issues that only affected a single screen.

Time saved: Manually running a WCAG check screen by screen across 14 screens and then synthesizing patterns by hand typically takes 3-4 hours. The chain compresses the review and grouping into roughly an hour of prompting and reading, leaving the rest for verification and the actual design fixes.

What to Do When It Breaks

• AI misses an issue a real contrast checker catches → Vision-based contrast estimates from a screenshot are approximate. Always run the actual hex values through a dedicated contrast checker for anything flagged as borderline, and treat the AI's severity ratings as a starting point, not a final ruling.
• Chain runs out of context partway through a large screen set → Split the audit into two sessions of roughly 10 screens each, and run Prompt 2 and Prompt 3 separately for each half, then manually merge the two prioritized lists.
• Findings feel repetitive or padded → Tighten Prompt 1 to skip minor cosmetic notes and focus only on WCAG-mapped criteria, so the model doesn't pad the list with subjective visual opinions.
• Real user data appears in a screenshot (a dashboard with live customer names or records) → Blur or replace real data with placeholder content before uploading any screen for review. Treat any screen pulled from a live environment the same way you'd treat a screenshot shared outside the company.

Variations

• Simpler version: Run just Prompt 1 on your 3-5 highest-traffic screens for a quick spot-check instead of a full-product pass.
• Extended version: Add a fourth prompt that drafts developer-ready ticket descriptions for each item in the prioritized fix list, so the audit output is ready to file directly into your issue tracker.

What to Do Next

• This week: Run the chain on one flow (not the whole product) to get a feel for how the pattern grouping works before committing to a full audit.
• This month: Schedule a recurring quarterly audit using this chain ahead of any major release, and compare each quarter's pattern list to see whether recurring issues are actually getting fixed.
• Advanced: Feed the prioritized fix list into the design-system-aware Custom GPT so proposed fixes automatically reference the correct existing component rather than a one-off patch.

Advanced guide for UX designer professionals. These techniques use more sophisticated AI features that may require paid subscriptions.