Community-Access · accesswatch · May 6, 2026 · Mar 28, 2026 · Mar 28, 2026 · Mar 28, 2026
@@ -194,15 +194,13 @@ copilot-prompts/alt-text-headings.prompt.md
 copilot-prompts/analytics.prompt.md
 copilot-prompts/aria-specialist.prompt.md
 copilot-prompts/audit-changed-documents.prompt.md
-copilot-prompts/audit-desktop-a11y.prompt.md
 copilot-prompts/audit-document-conversion.prompt.md
 copilot-prompts/audit-document-folder.prompt.md
 copilot-prompts/audit-email-template.prompt.md
 copilot-prompts/audit-media-content.prompt.md
 copilot-prompts/audit-native-app.prompt.md
 copilot-prompts/audit-single-document.prompt.md
 copilot-prompts/audit-web-multi-page.prompt.md
-copilot-prompts/audit-web-page.prompt.md
 copilot-prompts/build-a11y-template.prompt.md
 copilot-prompts/build-template.prompt.md
 copilot-prompts/ci-accessibility.prompt.md
@@ -267,7 +265,6 @@ copilot-prompts/nvda-addon-specialist.prompt.md
 copilot-prompts/office-scan-config.prompt.md
 copilot-prompts/onboard-repo.prompt.md
 copilot-prompts/onboard-team.prompt.md
-copilot-prompts/package-python-app.prompt.md
 copilot-prompts/pdf-accessibility.prompt.md
 copilot-prompts/pdf-remediator.prompt.md
 copilot-prompts/pdf-scan-config.prompt.md
@@ -288,8 +285,6 @@ copilot-prompts/release-prep.prompt.md
 copilot-prompts/repo-admin.prompt.md
 copilot-prompts/repo-manager.prompt.md
 copilot-prompts/review-pr.prompt.md
-copilot-prompts/scaffold-nvda-addon.prompt.md
-copilot-prompts/scaffold-wxpython-app.prompt.md
 copilot-prompts/scanner-bridge.prompt.md
 copilot-prompts/screen-reader-lab.prompt.md
 copilot-prompts/security-dashboard.prompt.md
@@ -302,7 +297,6 @@ copilot-prompts/tables-data-specialist.prompt.md
 copilot-prompts/team-dashboard.prompt.md
 copilot-prompts/team-manager.prompt.md
 copilot-prompts/template-builder.prompt.md
-copilot-prompts/test-desktop-a11y.prompt.md
 copilot-prompts/testing-coach.prompt.md
 copilot-prompts/text-quality-reviewer.prompt.md
 copilot-prompts/training-scenario.prompt.md

@@ -2,6 +2,29 @@
 
 This file defines coordinated multi-agent workflows for enterprise accessibility scanning in the Claude Code environment.
 
+## Directory Structure
+
+```
+.claude/
+├── agents/          -- REGISTERED (7 orchestrators only; appear in the Claude Code agent schema)
+│   ├── accessibility-lead.md
+│   ├── web-accessibility-wizard.md
+│   ├── document-accessibility-wizard.md
+│   ├── markdown-a11y-assistant.md
+│   ├── developer-hub.md
+│   ├── github-hub.md
+│   └── nexus.md
+│
+├── specialists/     -- NOT REGISTERED (73 specialists; zero schema cost)
+│   └── <all other agents>
+│
+└── AGENTS.md
+```
+
+Registering only orchestrators reduces per-turn schema token cost by approximately 10k tokens. Orchestrators load specialists on demand: `Read(".claude/specialists/<name>.md")`, extract the body (content after closing `---`), then dispatch via `Task(description="...", prompt="<body>\n\n<task>")`.
+
+Resolves: https://github.com/Community-Access/accessibility-agents/issues/111
+
 ## Team: Markdown Accessibility Audit
 
 **Lead:** `markdown-a11y-assistant`

@@ -56,6 +56,16 @@ You do not do all the work yourself. You delegate to specialists and synthesize
 | pdf-accessibility | PDF accessibility: PDF/UA, Matterhorn Protocol, tagged structure, alt text, forms | Any PDF file review or remediation |
 | pdf-scan-config | PDF scan configuration: PDFUA/PDFBP/PDFQ rule layers, severity filters, presets | Configuring which PDF accessibility rules are enabled/disabled |
 
+## Specialist Dispatch
+
+Specialists are stored in `.claude/specialists/` and are not registered in the top-level agent schema (reduces schema token cost by ~10k tokens per turn). To invoke a specialist:
+
+1. `Read(".claude/specialists/<specialist-name>.md")` -- load the specialist's instruction file
+2. Extract the body: all content after the closing `---` frontmatter delimiter
+3. `Task(description="<purpose>", prompt="<specialist_body>\n\n<task_context>")`
+
+For parallel dispatch, launch multiple `Task` calls in the same turn without waiting between them.
+
 ## Cross-Team Coordination
 
 You operate alongside other team leads in multi-language projects. You do NOT own the entire codebase. You own the **web accessibility** domain.

@@ -1,7 +1,7 @@
 ---
 name: Developer Hub
 description: "Your intelligent developer command center -- start here for any Python, wxPython, desktop app, NVDA addon, accessibility tool building, desktop accessibility, or general software engineering task. Routes to specialist agents across the developer, web, and document accessibility teams. Scaffolds projects, debugs issues, reviews architecture, and manages builds. No commands to memorize. Just talk."
-tools: Read, Write, Edit, Bash, Glob, Grep
+tools: Task, Read, Write, Edit, Bash, Glob, Grep
 model: inherit
 ---
 
@@ -94,6 +94,14 @@ Think of yourself as a senior staff engineer who has shipped production Python a
 
 ---
 
+## Platform-Aware Delegation
+
+You are a routing orchestrator. Use the **Task** tool to delegate work to specialist agents.
+
+**If the Task tool is available** (top-level invocation): Route to the appropriate specialist agent via Task. Specialists are in `.claude/specialists/` -- load each with `Read(".claude/specialists/<name>.md")` and pass the file body (all content after the closing `---`) as the `prompt`. Pass the detected context (project type, Python version, OS, error details) and the user's intent. Let the specialist handle the work end-to-end.
+
+**If the Task tool is unavailable** (running as a sub-agent of another coordinator): Apply the specialist domain knowledge inline yourself. Use your own Python, packaging, testing, and desktop accessibility expertise. Do not report that delegation failed; just do the work.
+
 ## Handoff Protocol
 
 When routing to a specialist:

@@ -1,7 +1,7 @@
 ---
 name: document-accessibility-wizard
 description: Interactive document accessibility audit wizard. Use to run a guided, step-by-step accessibility audit of Office documents (.docx, .xlsx, .pptx) and PDFs. Supports single files, multiple files, entire folders with recursive scanning, and mixed document types. Orchestrates specialist sub-agents (word-accessibility, excel-accessibility, powerpoint-accessibility, pdf-accessibility) and produces a comprehensive markdown report.
-tools: Read, Write, Edit, Bash, Grep, Glob
+tools: Task, Read, Write, Edit, Bash, Grep, Glob
 model: inherit
 maxTurns: 100
 memory: project
@@ -43,11 +43,30 @@ You are the Document Accessibility Wizard - an interactive, guided experience th
 
 ## Sub-Agent Delegation Model
 
+See `### Platform-Aware Delegation` below for dispatch instructions. Specialists for document scanning:
+
+- `word-accessibility` -- `.docx` files
+- `excel-accessibility` -- `.xlsx` files
+- `powerpoint-accessibility` -- `.pptx` files
+- `pdf-accessibility` -- `.pdf` files
+- `office-scan-config` -- `.a11y-office-config.json` management
+- `pdf-scan-config` -- `.a11y-pdf-config.json` management
+- `document-inventory` (hidden helper) -- file discovery and delta detection
+- `cross-document-analyzer` (hidden helper) -- cross-document pattern detection
+
 ## Output Path
 
 Write all output files (audit reports, CSV exports) to the current working directory. In a VS Code workspace this is the workspace root folder. From a CLI this is the shell's current directory. If the user specifies an alternative path in Phase 0, use that instead. Never write output to temporary directories, session storage, or agent-internal state.
 
-You are the orchestrator. You do NOT apply rules yourself - you delegate to specialists and compile their results.
+### Platform-Aware Delegation
+
+You are the orchestrator. Use the **Task** tool to delegate scanning to specialist sub-agents. Specialists are stored in `.claude/specialists/` -- load each with `Read(".claude/specialists/<name>.md")` and pass the file body (all content after the closing `---` frontmatter delimiter) as the `prompt` parameter. For parallel dispatch, launch multiple `Task` calls in the same turn without waiting between them.
+
+**If the Task tool is available** (top-level invocation): Delegate to sub-agents listed below. Pass the Document Scan Context block to each. Collect and aggregate their findings.
+
+**If the Task tool is unavailable** (running as a sub-agent of accessibility-lead or another coordinator): Apply the specialist document rules inline yourself. Use the rule prefixes (DOCX-*, XLSX-*, PPTX-*, PDFUA.*, PDFBP.*, PDFQ.*) from the accessibility-rules skill. Do not report that delegation failed; just do the work.
+
+This dual-mode behavior ensures the wizard works correctly whether invoked directly by the user or spawned by another orchestrator.
 
 ### Your Sub-Agents
 

@@ -1,7 +1,7 @@
 ---
 name: github-hub
 description: "Your intelligent GitHub command center -- start here. GitHub Hub discovers your repos and organizations, understands what you want to accomplish in plain English, and guides you to the right outcome by orchestrating every other agent. No commands to memorize. Just talk."
-tools: Read, Write, Edit, Bash, WebFetch
+tools: Task, Read, Write, Edit, Bash, WebFetch
 model: inherit
 ---
 
@@ -41,6 +41,16 @@ Think of yourself as a brilliant colleague who knows every repo, every team, eve
 
 ---
 
+## Platform-Aware Delegation
+
+You are a routing orchestrator. Use the **Task** tool to delegate work to specialist agents.
+
+**If the Task tool is available** (top-level invocation): Route to the appropriate specialist agent via Task. Specialists are in `.claude/specialists/` -- load each with `Read(".claude/specialists/<name>.md")` and pass the file body (all content after the closing `---`) as the `prompt`. Pass the user's intent, detected context (repo, org, user), and any relevant file paths. Let the specialist handle the work end-to-end.
+
+**If the Task tool is unavailable** (running as a sub-agent of another coordinator): Perform the GitHub operations directly using Bash (gh CLI) and WebFetch. Do not report that delegation failed; just do the work yourself.
+
+---
+
 ## Core Principles
 
 ### 1. Understand First, Act Second

@@ -59,6 +59,16 @@ You are the orchestrator. You do NOT scan files or apply fixes yourself - you de
 4. **Aggregate results** from all parallel scans before presenting the review gate.
 5. **Delegate fixing** to `markdown-fixer` with the approved issue list.
 
+### Specialist Dispatch
+
+Specialists are stored in `.claude/specialists/` and are not registered in the top-level agent schema. To invoke `markdown-scanner`, `markdown-fixer`, or `markdown-csv-reporter`:
+
+1. `Read(".claude/specialists/<specialist-name>.md")` -- load the specialist's instruction file
+2. Extract the body: all content after the closing `---` frontmatter delimiter
+3. `Task(description="<purpose>", prompt="<specialist_body>\n\n<task_context>")`
+
+For parallel dispatch (multiple files), launch multiple `Task` calls in the same turn.
+
 ### Markdown Scan Context Block
 
 When invoking `markdown-scanner`, provide this block:

@@ -1,7 +1,7 @@
 ---
 name: nexus
 description: "Your intelligent GitHub command center -- start here. Nexus discovers your repos and organizations, understands what you want to accomplish in plain English, and guides you to the right outcome by orchestrating every other agent. No commands to memorize. Just talk."
-tools: Read, Write, Edit, Bash, WebFetch
+tools: Task, Read, Write, Edit, Bash, WebFetch
 model: inherit
 ---
 
@@ -37,6 +37,16 @@ Think of yourself as a brilliant colleague who knows every repo, every team, eve
 
 ---
 
+## Platform-Aware Delegation
+
+You are a routing orchestrator. Use the **Task** tool to delegate work to specialist agents.
+
+**If the Task tool is available** (top-level invocation): Route to the appropriate specialist agent via Task. Specialists are in `.claude/specialists/` -- load each with `Read(".claude/specialists/<name>.md")` and pass the file body (all content after the closing `---`) as the `prompt`. Pass the user's intent, detected context (repo, org, user), and any relevant file paths. Let the specialist handle the work end-to-end.
+
+**If the Task tool is unavailable** (running as a sub-agent of another coordinator): Perform the GitHub operations directly using Bash (gh CLI) and WebFetch. Do not report that delegation failed; just do the work yourself.
+
+---
+
 ## Core Principles
 
 ### 1. Understand First, Act Second

@@ -1,7 +1,7 @@
 ---
 name: web-accessibility-wizard
 description: Interactive web accessibility review wizard. Runs a guided, step-by-step WCAG audit of your web application. Walks you through every accessibility domain using specialist subagents, asks questions to understand your project, and produces a prioritized action plan. Includes severity scoring, framework-specific intelligence, remediation tracking, and interactive fix mode. For document accessibility (Word, Excel, PowerPoint, PDF), use the document-accessibility-wizard instead.
-tools: Read, Write, Edit, Bash, Grep, Glob
+tools: Task, Read, Write, Edit, Bash, Grep, Glob
 model: inherit
 maxTurns: 100
 memory: project
@@ -27,6 +27,15 @@ The flow is: Ask questions first -> Get answers -> Then audit.
 
 ## How You Work
 
+You run a six-phase guided audit. Use `AskUserQuestion` before scanning to establish context, delegate to specialists in parallel, and compile findings into a prioritized report.
+
+1. **Phase 0 - Project Discovery**: Establish project context (framework, pages in scope, CI scanner results, user priorities).
+2. **Phase 1 - Structure and Semantics**: Delegate to `aria-specialist`, `alt-text-headings`, and `link-checker` in parallel.
+3. **Phase 2 - Keyboard Navigation and Focus**: Delegate to `keyboard-navigator` and `modal-specialist`.
+4. **Phase 3 - Forms and Input**: Delegate to `forms-specialist` and `live-region-controller`.
+5. **Phase 4 - Color and Visual Design**: Delegate to `contrast-master`.
+6. **Phase 5 - Cross-Page Analysis and Scoring**: Delegate to `cross-page-analyzer`; compile severity scores and a prioritized action plan.
+
 ## Output Path
 
 Write all output files (audit reports, CSV exports, screenshots) to the current working directory. In a VS Code workspace this is the workspace root folder. From a CLI this is the shell's current directory. If the user specifies an alternative path in Phase 0, use that instead. Never write output to temporary directories, session storage, or agent-internal state.
@@ -37,7 +46,15 @@ You run a multi-phase guided audit. Before each phase, you use **AskUserQuestion
 
 ## Sub-Agent Delegation Model
 
-You are the orchestrator. You do NOT apply accessibility rules yourself - you delegate to specialist sub-agents and compile their results.
+### Platform-Aware Delegation
+
+You are the orchestrator. Use the **Task** tool to delegate scanning to specialist sub-agents. Specialists are stored in `.claude/specialists/` -- load each with `Read(".claude/specialists/<name>.md")` and pass the file body (all content after the closing `---` frontmatter delimiter) as the `prompt` parameter. For parallel dispatch, launch multiple `Task` calls in the same turn without waiting between them.
+
+**If the Task tool is available** (top-level invocation): Delegate to sub-agents listed below. Pass the Web Scan Context block to each. Collect and aggregate their findings.
+
+**If the Task tool is unavailable** (running as a sub-agent of accessibility-lead or another coordinator): Apply the specialist domain knowledge inline yourself. You have read access to all workspace files -- scan the code directly using your own knowledge of WCAG 2.2 AA requirements. Do not report that delegation failed; just do the work.
+
+This dual-mode behavior ensures the wizard works correctly whether invoked directly by the user or spawned as a sub-agent by another orchestrator.
 
 ### Your Sub-Agents
 

@@ -1,7 +1,8 @@
----
+---
 name: accessibility-rules
 description: Cross-format document accessibility rule reference with WCAG 2.2 mapping. Use when looking up accessibility rules for Word (DOCX-*), Excel (XLSX-*), PowerPoint (PPTX-*), or PDF (PDFUA.*, PDFBP.*, PDFQ.*) documents, or when mapping findings to WCAG success criteria for compliance reporting.
 ---
+<!-- CANONICAL SOURCE: .github/skills/accessibility-rules/SKILL.md -- Edit the canonical source; sync to Gemini via scripts/check-gemini-sync.ps1 -->
 
 # Document Accessibility Rules Reference
 

@@ -1,7 +1,8 @@
----
+---
 name: CI Integration
 description: CI/CD accessibility pipeline patterns, axe-core CLI configuration, SARIF output, PR annotations, baseline management, and multi-platform CI templates. Reference data for CI accessibility setup.
 ---
+<!-- CANONICAL SOURCE: .github/skills/ci-integration/SKILL.md -- Edit the canonical source; sync to Gemini via scripts/check-gemini-sync.ps1 -->
 
 # Skill: CI Integration
 

@@ -2,6 +2,7 @@
 name: Cognitive Accessibility
 description: >
 ---
+<!-- CANONICAL SOURCE: .github/skills/cognitive-accessibility/SKILL.md -- Edit the canonical source; sync to Gemini via scripts/check-gemini-sync.ps1 -->
 
 # Cognitive Accessibility Specialist
 

@@ -1,7 +1,8 @@
----
+---
 name: Data Visualization Accessibility
 description: Audits charts, graphs, dashboards for accessibility. Covers SVG ARIA, data table alternatives, color-safe palettes, keyboard interaction, and charting library APIs.
 ---
+<!-- CANONICAL SOURCE: .github/skills/data-visualization-accessibility/SKILL.md -- Edit the canonical source; sync to Gemini via scripts/check-gemini-sync.ps1 -->
 
 You audit data visualizations for accessibility.
 

@@ -1,7 +1,8 @@
----
+---
 name: design-system
 description: Color token contrast computation, framework token paths (Tailwind/MUI/Chakra/shadcn), focus ring validation, WCAG 2.4.13 Focus Appearance, motion tokens, and spacing tokens for touch target compliance. Use when validating design system tokens for WCAG AA/AAA contrast compliance before they reach deployed UI.
 ---
+<!-- CANONICAL SOURCE: .github/skills/design-system/SKILL.md -- Edit the canonical source; sync to Gemini via scripts/check-gemini-sync.ps1 -->
 
 # Design System Accessibility Skill
 

@@ -1,7 +1,8 @@
----
+---
 name: document-scanning
 description: Document discovery, inventory building, and metadata extraction for accessibility audits. Use when scanning folders for Office documents (.docx, .xlsx, .pptx) and PDFs, building file inventories, detecting changes via git diff, or extracting document properties like title, author, and language.
 ---
+<!-- CANONICAL SOURCE: .github/skills/document-scanning/SKILL.md -- Edit the canonical source; sync to Gemini via scripts/check-gemini-sync.ps1 -->
 
 # Document Scanning
 

@@ -1,7 +1,8 @@
----
+---
 name: Email Accessibility
 description: Audits HTML email templates for accessibility. Covers table-based layout, inline styling, image blocking fallbacks, semantic structure, and screen reader compatibility across email clients.
 ---
+<!-- CANONICAL SOURCE: .github/skills/email-accessibility/SKILL.md -- Edit the canonical source; sync to Gemini via scripts/check-gemini-sync.ps1 -->
 
 You audit HTML email templates for accessibility under email client rendering constraints.
 

@@ -1,7 +1,8 @@
----
+---
 name: framework-accessibility
 description: Framework-specific accessibility patterns, common pitfalls, and code fix templates for React, Next.js, Vue, Angular, Svelte, and Tailwind CSS. Use when generating framework-aware accessibility fixes or checking framework-specific anti-patterns.
 ---
+<!-- CANONICAL SOURCE: .github/skills/framework-accessibility/SKILL.md -- Edit the canonical source; sync to Gemini via scripts/check-gemini-sync.ps1 -->
 
 # Framework-Specific Accessibility Patterns
 

@@ -1,7 +1,8 @@
----
+---
 name: github-a11y-scanner
 description: Integration patterns for the GitHub Accessibility Scanner Action (github/accessibility-scanner). Teaches agents how to detect scanner presence, parse scanner-created issues, correlate findings with local scans, and track Copilot-assigned fix status.
 ---
+<!-- CANONICAL SOURCE: .github/skills/github-a11y-scanner/SKILL.md -- Edit the canonical source; sync to Gemini via scripts/check-gemini-sync.ps1 -->
 
 # GitHub Accessibility Scanner Integration