Accessibility Fix Desk: agent fixes for common WCAG 2.2 failures, checked against the original, owner approved
by PaperChaseWebb
Let an agent fix the accessibility failures found on almost every site: missing alt text, unnamed links and buttons, unlabeled fields, faint text, blocked zoom, hidden focus. It writes fixed copies, never the live page. A fix check compares each copy with the original and refuses deleted or hidden content, new scripts and overlays, alt text you never wrote, colors outside your palette and aria-labels that drop the visible words. A claims check stops "fully compliant" statements. You approve at your terminal and put the page live yourself.
- Price
- $29
- Difficulty
- Intermediate
- Setup
- 1 hr
- Format
- .md file
accessibility-fix-desk-agent-fixes-checked-against-original-owner-approved.md · 201 KB · instant download after checkout
Want this built for you instead? We can build it for you.
What you end up with
Your pages lose the common, checkable accessibility failures one approved fix at a time, with nothing deleted or hidden, no overlay, and a statement that only says what was checked.
What you get
- 21 sections and an appendix in one .md file, with a one-line unpack for all 68 files
- a11y_check.py scan: 20 checks for common WCAG 2.2 A and AA failures, including contrast worked out from the page's own CSS
- a11y_check.py fix: 15 checks that compare the agent's copy with the original, so fixing by deleting, hiding or adding a script fails
- Alt text from your own image list and colors from your own palette, with anything undecided left for you
- a11y_check.py claims: stops fully compliant, automatic, lawsuit and no issues claims in statements and reports
- The owner step: approve.py needs a real terminal and part of the fix hash, refuses agent sessions and piped input, and checks the live page has not changed since the fix
- A 10-page eval set, rules and two skills for Grok Bot, Claude Code, Codex and Hermes, six prompts and five templates
- The test harness: 80 unit tests and 518 planted faults you can repeat
Who it is for
Site owners, web teams and agencies who want an agent to clear the common accessibility failures page by page, with a person approving every change and doing the hand tests a program cannot.
Works with
Solid chips: the agent is named in this listing's own tools list. Dashed chip: Plain Markdown, so it also pastes into other agents such as Grok Bot, Hermes, OpenClaw or Muse. We list those as compatible by format only, not as tested.
Tools needed
Cost to run: The kit is the only purchase. Everything runs locally with Python; the checker makes no network calls, logs in to nothing and changes no website.
Time to result: The worked example in 5 minutes; your first three pages scanned and fixed in about an hour
How it works
Unpack and test
Save the kit as pack.md in an empty folder, run the unpack line from Appendix A, then sh test_all.sh.
Expected: 68 files written; 80 tests OK; 518 of 518 planted faults caught.
Watch the flow
Run sh example/run_example.sh and read it next to section 17.
Expected: A bike shop home page with 12 findings, an agent's first-try fix with 14, the second fix clean, a first-try statement with 7 findings, the free check finding 8 and a piped approval refused.
Put in your site
Fill in site/site.json and site/images.json from the templates, save your pages into site/pages, then run python3 desk/a11y_check.py site.
Expected: No findings, or a short list of settings to fix first.
Scan and fix three pages
Run a11y_check.py scan site/pages, have your agent follow skills/a11y-fix for three pages, then a11y_check.py fix queue/fixes.
Expected: Fixed copies in queue/fixes with 0 high findings, and a note listing what is left for you and what to test by hand.
Sample prompts, free to use
Follow skills/a11y-fix for the three pages in site/pages with the most high findings. Never edit the site. List each fixed page with the codes fixed and the codes left for me.
Prompt 1 of 6.
List every image in site/pages that is not in images.json, with the page, the file name and the words around it. Do not write alt text and do not edit images.json.
Prompt 3 of 6.
Follow skills/statement-check for the pages fixed this month and write it to reports/. Run the claims check and fix every finding.
Prompt 5 of 6.
Free preview of the file
# Accessibility Fix Desk: agent fixes for common WCAG 2.2 failures, checked against the original, owner approved Version 1.0, 2026-11-01. By PaperChaseWebb. For site owners, small businesses, web teams and agencies who want an AI agent to fix the accessibility failures that show up on almost every site, without deleting content, hiding it, inventing alt text, adding an overlay or claiming the site is compliant. Works with Grok Bot, Claude Code, Codex and Hermes. Python 3.8 or newer, standard library only, no API keys, no logins. The kit and the agent never edit your website. Tested with 80 unit tests and 518 planted faults on Linux with Python 3.13, 3.12, 3.11, 3.9 and 3.8. All business and page data in this kit is made up. The same few failures cover most of what an automated check finds: text too faint to read, images with no text, links and buttons with no name, form fields with no label, no page language, zoom turned off, keyboard focus made invisible. In the WebAIM Million 2026 study, 95.9% of home pages had at least one detected failure, and low contrast text alone was on 83.9% of them. These are small fixes, and an AI agent can make them fast. It also gets them wrong in ways that are hard to spot by eye. It deletes the icon link it could not name. It wraps the faint text in aria-hidden so the contrast warning goes away. It writes "best bike repair in town, award-winning bike shop" as alt text. It sets an aria-label of "Open booking menu" on a button that says "Book
Test record
- Date
- Nov 1, 2026
- What was run
- tests/test_scan.py (24), test_contrast.py (10), test_fix.py (20), test_modes.py (14), test_approve.py (8, on a real pseudo terminal) and test_mini.py (4); tests/mutate.py with 408 planted faults across page problems, bad fixes, statements, agent rules, site settings and approve.py, plus 110 against a11y_mini.py; a11y_check.py set on the 10-page eval set; example/run_example.sh end to end; and the unpack line with test_all.sh. Linux, Python 3.13, 3.12, 3.11, 3.9 and 3.8.
- Result
- 80 passed, 0 failed on all five versions; 518 of 518 planted faults caught and 11 of 11 clean baselines clean; the eval set had 0 findings; the example caught the first-try fix with 14 findings and the first-try statement with 7, and refused piped approval; the unpack line wrote every file and test_all.sh passed there.
- Not run
- A live Grok Bot, Claude Code, Codex or Hermes session. Screen readers, keyboard and zoom on real devices. Pages built by JavaScript. Non-English pages. macOS or Windows.
Copied from this item's dated test notes. What was not run stays in.
Worth knowing
- The agent rules were checked by the checker but not loaded inside a live Grok Bot, Claude Code, Codex or Hermes session for this release. The eval fixes were written by hand to the rules. Give your agent the entrance exam in section 15 before you rely on it.
- This is not a full WCAG audit and not legal advice. It checks about 20 common failures that a program can find; many WCAG criteria need a person. Nothing in the kit says a site meets any law or standard.
- The checker reads the HTML you export and simple CSS in the page. It does not run scripts or render the page, so external stylesheets, layout and pages built in the browser need hand checks.
- Screen readers, keyboard use and zoom on real devices were not tested by the kit. Every fix note lists those hand tests for the owner.
- The owner step assumes the agent cannot run approve.py at a terminal. Run it as a separate user; agent rules alone are not a hard boundary.
- All businesses and pages in the kit are made up. English only. Tested on Linux only; macOS and Windows were not tested.
- On 2026-11-01 on Linux: 80 unit tests and 518 of 518 planted faults passed on Python 3.13, 3.12, 3.11, 3.9 and 3.8; the 10-page eval set was clean; the unpack line wrote every file and test_all.sh passed in that folder.
Questions
Which agents does it work with?
Any agent that can read files and run a script. The kit has rules for AGENTS.md and CLAUDE.md, Claude Code and Codex notes, Hermes notes and a Grok Bot standing order. Those were checked by the checker, not loaded in each agent for this release.
Will it change my website?
No. The agent writes fixed copies into a queue folder. You approve each one at a terminal and put it on your site yourself.
Will it make my site compliant?
No tool can promise that, and this kit will not say it. It finds and fixes common failures a program can check, records what was done, and lists what a person still has to test.
Why not use an overlay widget?
An overlay adds a script on top of the page instead of fixing the markup, and the FTC fined one vendor $1 million in 2025 over its compliance claims. The scan flags overlays and the fix check refuses to add one.
Can you set it up for us?
Yes. Our Install sprint sets up your palette, image list, eval set and agent, or Managed Ops runs the scans, fixes and statement for you each month. Book a call or write to [email protected].
Details
Version 1.0 · Updated Nov 1, 2026 · Markdown version of this page
License: Use inside your own business or on your clients' sites. Do not resell or republish the kit itself. Download terms
Related
Accessibility Fix Desk: safe fix rules and a ten-check page scan (free sample)
From Accessibility Fix Desk: the mistakes agents make when they fix accessibility, eight rules to paste into your agent, and a read-only script that runs ten checks on one HTML page, from language, title, alt text and unnamed links and buttons to unlabeled fields, blocked zoom, tabindex, removed focus outlines and overlay scripts.
Search Console Fix Desk: better page titles and descriptions from your Search Console data, checked, owner approved, measured
Let an agent find the pages Google shows but people skip, and draft better titles and meta descriptions without inventing claims or touching your site. A planner compares each page with your own click rates at the same positions and decides rewrite, owner, hold or skip. A checker refuses drafts with claims or numbers the page never states, keyword stuffing, ranking promises, competitor names, noindex or redirect ideas and script overlays. You approve at your terminal and make the edit, and measure compares before and after against pages you did not touch.
Support Agent from Your Docs: grounded reply drafts, owner approval and a checker
Turn your help docs into a support agent that only says what your docs say. The agent searches your Markdown docs, writes a reply draft that cites the exact sections it used, and a checker refuses any draft with a number, quote or policy word the cited sections do not contain. Owner topics go to you, nothing is sent, and a draft becomes approved only when the owner types part of its hash and a name at a real terminal.
Review Reply Desk: AI-drafted review replies, checked for privacy and offers, approved by the owner
Let an agent draft your Google and Yelp review replies without the mistakes that cost money. A checker refuses drafts that use a last name or details from your records, offer a discount for a better review, ask for a review to be changed, argue, invent numbers or promises, or repeat your last reply word for word. Health topics, allergies and legal threats go to the owner, clinics get a stricter privacy mode, and nothing is posted until the owner approves it at a real terminal.
Agency Open-Loop System: meeting notes to tasks, client follow-ups and a weekly open-loop report
A Markdown operating system plus a small checker for owner-led agencies: turn every meeting into commitments with an owner, a date, a quote and a done test, draft client follow-ups as approval cards you answer by number, and get one weekly report of what is still open. Includes AGENTS.md with six owner stops, three skills, templates and loops_check.py.
Ask-first rules for AI agents: hard nos, approval cards and a test set
Paste-in rules that make an agent ask before it sends, spends, posts or deletes, define what a real yes is, and stop retries after a no, plus 12 test scenarios to check it.
Content System Kit: run a content engine with AI agents and keep a human on the publish button
A prompt pack and method for founders, small teams and agencies. Audience filters, one owned format, a 90-post test with review rules set in advance, a weekly batch loop with one approval gate, an editor queue with a rights check, transcript-first cleanup, a packaging watch on public data only, teardown without replication, a craft seat map, a cadence plan, an engagement honesty audit, a failure table, a worked example, a rules block and 14 prompts. Includes content_check.py, AGENTS.md with six owner stops, two skills and 15 templates.