microsoft/create-site
>- Creates a new Power Pages code site (SPA) using React, Angular, Vue, or Astro. Guides through component planning, design, implementation, validation, and deployment. Use when the user wants to create, build, or scaffold a new Power Pages website or portal.
npx skills add https://github.com/microsoft/power-platform-skills --skill create-site
> Plugin check: Run node "${PLUGIN_ROOT}/scripts/check-version.js" — if it outputs a message, show it to the user before proceeding.
Guide the user through creating a complete, production-quality Power Pages code site from initial concept to deployed site. Follow a systematic approach: discover requirements, scaffold and launch immediately, plan components and design, implement with design applied, validate, review, and deploy.
browser_navigate + browser_snapshot) to verify every significant change. Do NOT take screenshots — only use accessibility snapshots to check page structure and content.public/scaffold-status.json. Update this file before every AskUserQuestion (to raise the "waiting for your input" banner so the user doesn't miss a terminal prompt) and before each implementation step in Phase 5 (so the progress-bar label matches what you're actually doing while the decorative spinner continues its default cycle). See Live Preview Status Protocol.https://images.unsplash.com/photo-{id}?w={width}&h={height}&fit=crop URLs with specific photo IDs found via WebSearch. Never leave image placeholders or broken <img> tags pointing to nonexistent files.Constraint: Only static SPA frameworks are supported (React, Vue, Angular, Astro). NOT supported: Next.js, Nuxt.js, Remix, SvelteKit, Liquid.
Initial request: $ARGUMENTS
<!-- not-a-gate: prose-only section — mentions of AskUserQuestion here describe the live-status protocol that wraps every real prompt in Phases 3/4/8; the actual gates are catalogued in §6.13 of references/approval-gates.md and marked at their call sites below -->
While the scaffold loading screen is visible (from Phase 2.6 until the Home page itself is replaced in Phase 5), the loader polls GET /scaffold-status.json every 1.5 seconds. The message you write into <PROJECT_ROOT>/public/scaffold-status.json appears as the label under the progress bar, and awaitingInput controls the "waiting for your input" banner. The decorative spinner above the progress bar continues its built-in phrase cycle; keep the progress-bar label current so the loader still reflects what is actually happening.
Why this matters: When the browser with the loader takes over the user's screen, a prompt in the terminal can sit unanswered for a long time because the user doesn't realize anything is waiting. The banner makes it obvious.
File shape (all fields optional — omit any field you don't want to change):
{
"message": "Creating Contact page",
"awaitingInput": false,
"inputPrompt": "Please check your terminal to respond."
}
message — one short present-participle phrase shown as the status line under the progress bar in the loader (replacing the default "Getting started…" / "Setting up infrastructure…" cycle). Include the grouping context inline when it helps (e.g., "Creating Footer component (shared components)").awaitingInput — when true, a prominent pulsing banner appears at the top of the loader and stays visible until this field is cleared. Set this before every AskUserQuestion call and clear it (false) immediately after the user answers.inputPrompt — short context for the banner (e.g., "Choose a framework"). Optional.When to update the file:
{ "message": "Planning your site", "awaitingInput": false }.AskUserQuestion that runs while the scaffold is visible (Phases 3, 4, and any in-scaffold prompt in Phase 5): set awaitingInput: true with a short inputPrompt. After the user answers, write again with awaitingInput: false.message to the specific action. Examples: "Applying design tokens", "Creating Navbar component", "Creating Contact page".public/scaffold-status.json so it isn't deployed with the site.Write the file with the Write tool (atomic overwrite). You do not need to read it first.
Goal: Understand what site needs to be built and what problem it solves
Actions:
<!-- gate: create-site:1.purpose | category=plan | cancel-leaves=nothing -->
> 🚦 Gate (plan · create-site:1.purpose): Multi-question prompt collecting site name, framework, purpose, audience, and target directory. Determines what gets scaffolded. Fires only on the "site purpose unclear" branch (step 3 below).
>
> Trigger: Phase 1 when site purpose was not provided in $ARGUMENTS.
> Why we ask: Wrong framework picked → wrong template copied into the wrong directory; cleanup is annoying.
> Cancel leaves: Nothing — no scaffolding has started yet.
AskUserQuestion:| Question | Header | Options |
|----------|--------|---------|
| What should the site be called? (e.g., "Contoso Portal", "HR Dashboard") | Site Name | *(free text — use a single generic option so the user types a custom name via "Other")* |
| Which frontend framework? | Framework | React (Recommended), Vue, Angular, Astro |
| What is the site's purpose? | Purpose | Company Portal, Blog/Content, Dashboard, Landing Page |
| Who is the target audience? | Audience | Internal (employees, partners), External (public-facing customers) |
| Where should the project be created? | Location | Current directory, New folder in current directory (Recommended), Any other directory |
<cwd>.__SITE_NAME__ inside the cwd. Project root = <cwd>/__SITE_NAME__/.After resolving, confirm: "The site will be created at <resolved path>."
Store this as PROJECT_ROOT.
__SITE_NAME__ (Title Case, e.g., Contoso Portal)__SITE_SLUG__ (kebab-case derived from site name, e.g., contoso-portal)__SITE_DESCRIPTION__ (one-line description based on name + purpose)Audience influences site generation:
Output: Clear statement of site purpose, framework, audience, derived naming values, and project location
Goal: Get a running site immediately so the user has something to preview while features and design are planned
> The scaffold is a temporary branded loading screen — it shows a Power Pages animated "Building your site" experience with orbiting elements, status messages, and feature cards. Its only purpose is to get the dev server running quickly so the user has something to look at while you plan and build. During Phase 5 (Implementation), the entire scaffold — including theme.css, Layout, Home page, and all placeholder components — is completely replaced with the user's actual site: their chosen typography, color palette, pages, components, and navigation. Do NOT try to build on top of the loading screen; replace it entirely.
> See ${PLUGIN_ROOT}/references/framework-conventions.md for the full framework → build tool → router → output path mapping.
Actions:
> ${PLUGIN_ROOT} is already resolved to the plugin's absolute path at runtime. Use it directly in Glob/Read paths — do NOT search for the plugin directory.
Read and copy all files from the matching asset template to the project directory:
| Framework | Asset Directory |
|-----------|----------------|
| React | ${PLUGIN_ROOT}/skills/create-site/assets/react/ |
| Vue | ${PLUGIN_ROOT}/skills/create-site/assets/vue/ |
| Angular | ${PLUGIN_ROOT}/skills/create-site/assets/angular/ |
| Astro | ${PLUGIN_ROOT}/skills/create-site/assets/astro/ |
Use Glob to discover all files in the asset directory, Read each file, then Write to the project directory preserving the relative path structure.
Also copy the shared loader icon that the scaffold references from its CSS (url('/power-pages-icon.png')):
Read the binary file ${PLUGIN_ROOT}/skills/create-site/assets/shared/power-pages-icon.png and Write it to <PROJECT_ROOT>/public/power-pages-icon.png. (All four supported frameworks serve public/ at the web root, so the same /power-pages-icon.png URL works for every framework.)
Seed the live status file so the loader shows a real message the moment it mounts. Write <PROJECT_ROOT>/public/scaffold-status.json:
{ "message": "Planning your site", "awaitingInput": false }
See Live Preview Status Protocol for the full contract — from here on, update this file before every AskUserQuestion and before each Phase 5 implementation step.
After copying, replace all __PLACEHOLDER__ tokens in every file. Use Edit with replace_all: true on each file.
__SITE_NAME__, __SITE_SLUG__, __SITE_DESCRIPTION__).> Note: The scaffold loading screen uses hardcoded Power Pages branding colors — there are no color placeholders (__PRIMARY_COLOR__, etc.) to replace. The user's chosen color palette is applied fresh during Phase 5 when the scaffold is completely replaced.
Rename gitignore → .gitignore in the project root (stored without dot prefix to avoid git interference in the plugin repo).
Run npm install before initializing git so that package-lock.json is included in the initial commit:
cd "<PROJECT_ROOT>"
npm install
> Astro requires Node 22.12 or newer.
> Astro 7 exits with Node.js vX is not supported by Astro! on anything older, so when the chosen framework is Astro, run node --version first and ask the user to upgrade before continuing.
Initialize a git repo and make the first commit. This captures all template files AND package-lock.json in one clean baseline:
cd "<PROJECT_ROOT>"
git init
git add -A
git commit -m "Initial scaffold: __SITE_NAME__ (__FRAMEWORK__)"
From this point, commit after every significant milestone so any breaking change can be reverted.
This MUST happen now — before any planning or customization begins. The dev server gives the user a live preview while features and design are being planned:
cd "<PROJECT_ROOT>"
npm run dev
Run npm run dev in the background using Bash with run_in_background: true. Note the local URL (typically http://localhost:5173 for Vite or http://localhost:4200 for Angular or http://localhost:4321 for Astro).
Immediately after the dev server starts, verify the scaffold is working:
mcp__plugin_power-pages_playwright__browser_navigate to open the dev server URLmcp__plugin_power-pages_playwright__browser_snapshot to verify the page loaded correctly (do NOT take screenshots — only use accessibility snapshots)http://localhost:5173 — open it in your browser to follow along as I build.")> GATE: Do NOT proceed to Phase 3 until ALL of the following are true:
>
> 1. Template files copied and placeholders replaced
> 2. Git repo initialized with initial scaffold commit
> 3. npm install completed successfully
> 4. Dev server is running in the background (npm run dev)
> 5. Playwright has opened the site and verified it loads via browser_snapshot
> 6. The dev server URL has been shared with the user
>
> If any of these are not done, complete them now before moving on.
Output: Running dev server with verified scaffold, URL shared with user
Goal: Determine what pages, components, and design elements the site needs — while the user previews the running scaffold
<!-- gate: create-site:3.requirements | category=plan | cancel-leaves=nothing -->
> 🚦 Gate (plan · create-site:3.requirements): Three sub-prompts (features multi-select, aesthetic, mood) — shape the Phase 4 plan and the Phase 5 implementation. Fires at step 2 of the action list below.
>
> Trigger: Phase 3 entry; scaffold loader is up.
> Why we ask: Wrong feature set / aesthetic gets baked into the rendered plan — the Phase 4.7 gate would still catch most errors, but it's wasteful to defer the catch.
> Cancel leaves: Nothing — scaffold loader files are throwaway artifacts replaced wholesale in Phase 5.
Actions:
Write <PROJECT_ROOT>/public/scaffold-status.json: { "message": "Planning your site", "awaitingInput": true, "inputPrompt": "Features, aesthetic, and mood — please answer in the terminal." }
Immediately after the user answers, Write the same file again with "awaitingInput": false so the banner disappears.
AskUserQuestion to collect feature and design requirements:| Question | Header | Options |
|----------|--------|---------|
| Which features? (multi-select) | Features | *(generate 3-4 context-aware options based on the site name, purpose, and audience from Phase 1)* |
| What aesthetic direction do you want? | Aesthetic | Minimal & Clean (Recommended), Bold & Vibrant, Dark & Moody, Warm & Organic |
| What's the overall mood? | Mood | Professional & Trustworthy (Recommended), Creative & Playful, Technical & Precise, Elegant & Premium |
> Feature options are NOT hardcoded. Infer relevant features from Phase 1 answers. For example:
> - "HR Dashboard" + Internal → Employee Directory, Leave Requests, Announcements, Org Chart
> - "Contoso Portal" + External → Contact Form, Service Catalog, Knowledge Base, FAQ
> - "Partner Hub" + Internal → Document Library, Partner Directory, Deal Tracker, Notifications
>
> Always generate options that make sense for the specific site — never reuse a fixed list.
>
> If you include an Authentication feature option, describe it generically as "Login/signup for tracking application status" or similar. Do NOT mention a specific identity provider (e.g., "Entra ID", "SAML", "Google") in the feature description — the /power-pages:setup-auth skill will ask the user which provider they want.
/add-ai-webapi populates the slots later when the user is ready. Use AskUserQuestion with multi-select to let the user opt in:| Question | Header | Options |
|----------|--------|---------|
| Which AI summarization features should the site have? (multi-select — each can be added later with /add-ai-webapi) | AI Summaries | *(generate 2-4 context-aware options plus "None for now")* |
> Options are NOT hardcoded. Infer relevant AI summary features from Phase 1 and the features picked above. Examples:
> - "HR Dashboard" + Leave Requests feature → "Data summarization for leave requests", "Search summary on the knowledge base"
> - "Contoso Portal" + Knowledge Base → "Search summary on site-wide search", "Data summarization for articles"
> - "Customer Self-Service" + Support Cases / Incidents → "Data summarization for support cases (Microsoft-shipped recipe)", "Data summarization for attached knowledge articles"
>
> Treat the standard incident table like any other Dataverse table — propose Data
> Summarization for it when the site handles support cases, but don't force the Microsoft-shipped
> recipe ($select=description,title + the portal-comments expand) unless that genuinely fits
> the user's UX. A custom case-like table or a different facet of the standard incident is a
> regular Data Summarization pick. Always include None for now so the user can defer. Do
> NOT integrate the APIs in this skill — only record the user's picks so Phase 4's plan can
> mention them and Phase 8 can suggest /add-ai-webapi as a recommended next step.
>
> Capture the selection in memory as AI_SUMMARY_PICKS — a list of one or more of: search-summary, data-summarization.
AI_SUMMARY_PICKS, decide which page will carry the AI surface and store the mapping as AI_SUMMARY_PLACEMENTS. This is what Phase 4 shows the user and what Phase 5 reserves slots for. Use the feature selection from step 2 — and treat this mapping as an *input* to the page list Claude proposes in step 7: if a pick has no natural target page, add one to the plan so the summary has a home:| Pick | Default target page | If no matching page is planned |
|------|--------------------|-------------------------------|
| search-summary | A search / search-results page (e.g., SearchResults, Search) | Add a search page to the plan so the summary has a home |
| data-summarization | The detail page of the table the user called out (e.g., ProductDetail for products, CaseDetail for support cases) — ask the user if ambiguous | Propose adding a detail page; if rejected, fall back to a list/dashboard page |
AI_SUMMARY_PLACEMENTS shape: one record per placement, e.g.
[{ pick: "data-summarization", targetPage: "CaseDetail", marker: "POWERPAGES:AI-SLOT kind=data-summarization" }].
The marker string is the comment tag Phase 5 emits into the page source as a reserved anchor that /add-ai-webapi later finds. Keep the shape uniform — one marker per placement, always the same tag, so the follow-up skill's explore step can grep for them deterministically.
${PLUGIN_ROOT}/skills/create-site/references/design-aesthetics.mdAI_SUMMARY_PLACEMENTS from step 4 implies a page that wasn't already in the plan (e.g., a CaseDetail page for a data-summarization pick on the support-case table), add it to the page list now. Present the component plan to the user as a table: | Component Type | Count | Details |
|---------------------|-------|---------|
| Pages | 4 | Home, About, Services, Contact |
| Shared Components | 3 | Navbar, Footer, ContactForm |
| Design Elements | 4 | Google Fonts (Playfair Display + Source Sans Pro), Color palette (6 CSS vars), Page transitions, Gradient backgrounds |
| Routes | 4 | /, /about, /services, /contact |
theme.css during Implementation (Phase 5) when the scaffold loading screen is completely replaced:| CSS Variable | Description | Value |
|-------------|-------------|-------|
| --color-primary | Primary hex color | *(choose based on aesthetic + mood)* |
| --color-secondary | Complementary hex color | *(choose based on aesthetic + mood)* |
| --color-bg | Background color | *(choose based on aesthetic + mood)* |
| --color-surface | Surface/card color | *(choose based on aesthetic + mood)* |
| --color-text | Main text color | *(choose based on aesthetic + mood)* |
| --color-text-muted | Muted text color | *(choose based on aesthetic + mood)* |
Output: Confirmed list of pages, components, design elements, and routes to create
Goal: Render the implementation plan as an HTML document, open it in the user's default browser, and get approval before starting implementation.
> Why HTML instead of a chat message: A structured HTML plan (like the ones produced by /integrate-backend, /add-server-logic, and /add-cloud-flow) lets the user skim sections, compare swatches, and preview typography — all impossible in a terminal. The scaffold loader in their browser may also be full-screen, so surfacing the plan in a new tab puts it where they can actually read it.
Read the design aesthetics reference: ${PLUGIN_ROOT}/skills/create-site/references/design-aesthetics.md. Every field you populate below should be justified by the chosen aesthetic + mood from Phase 3.
> AI Readiness in the plan. If AI_SUMMARY_PLACEMENTS from Phase 3 is non-empty, reflect each placement in the matching PAGES_DATA entry's description or content — e.g., *"Reserved slot for an AI summary card; populated later by /add-ai-webapi. The page ships without AI."* This keeps the user's expectation honest: the site does not depend on generative-AI features being enabled on the tenant, and there is no "Run /add-ai-webapi" placeholder visible to end-users. If AI_SUMMARY_PLACEMENTS is empty, omit any AI references from the plan.
Assemble a single JSON object with the following keys. The plan template rejects any data that's missing a required key, so include all of them.
| Key | Type | Content |
|-----|------|---------|
| SITE_NAME | string | Title-case site name from Phase 1 |
| PLAN_TITLE | string | Always "Implementation Plan" |
| FRAMEWORK | string | React / Vue / Angular / Astro |
| AESTHETIC | string | Chosen aesthetic (e.g., Minimal & Clean) |
| MOOD | string | Chosen mood (e.g., Professional & Trustworthy) |
| SUMMARY | string | One paragraph describing what the site is and who it serves |
| TYPOGRAPHY_DATA | object | { primary: { name, sample, reason }, secondary: { name, sample, reason } } — name must be a real Google Font family |
| PALETTE_DATA | array | [{ var, hex, description }] — one entry per CSS variable (primary, secondary, bg, surface, text, text-muted) |
| MOTION_DATA | array | [{ label, description }] — page transitions, hover states, etc. |
| BACKGROUNDS_DATA | array | [{ label, description }] — hero backgrounds, section treatments, patterns |
| PAGES_DATA | array | [{ name, route, description, content: [...], components: [...] }] — content is an outline of what's on the page, components is shared component names used |
| COMPONENTS_DATA | array | [{ name, purpose, usedBy: [...] }] — shared components with the page names that consume them |
| ROUTES_DATA | array | [{ path, page }] — every route the router will register |
| REVIEW_DATA | array of strings | Verification checklist items (e.g., "All pages load without console errors") |
| DEPLOYMENT_DATA | array | [{ title, description, recommended?: boolean }] — mark exactly one as recommended: true |
Write the data for the user, not for internal tooling — phrase description and reason fields in plain language.
Pick an output path under <PROJECT_ROOT>/docs/. Default is create-site-plan.html; if that file already exists, pick a descriptive variant like create-site-plan-v2.html (the render script refuses to overwrite existing files).
node "${PLUGIN_ROOT}/scripts/render-createsite-plan.js" --output "<PROJECT_ROOT>/docs/create-site-plan.html" --data-inline '<json-string>'
Use --data-inline so no temp JSON file is written. If the JSON is too large for a single shell argument, write it to a temp file and use --data <path> instead, then delete the temp file after the render succeeds.
The script prints {"status":"ok","output":"<path>"} on success. Capture and use that actual output path for the next step.
Open <OUTPUT_PATH> in the default browser using the platform-appropriate file opener for the current environment. For example, use open on macOS, xdg-open on Linux, or the equivalent default-browser opener available on Windows.
Keep the terminal message short — the full plan lives in the HTML file now. Include:
Do NOT dump the full plan contents into the terminal — that defeats the purpose of the HTML view.
The user may still be looking at the full-screen scaffold loader when you ask for approval. Write <PROJECT_ROOT>/public/scaffold-status.json:
{ "message": "Ready to build", "awaitingInput": true, "inputPrompt": "Plan approval needed — review the plan in your browser and answer in the terminal." }
Immediately after the user answers, Write the same file again with "awaitingInput": false.
<!-- gate: create-site:4.7.plan-approval | category=plan | cancel-leaves=nothing -->
> 🚦 Gate (plan · create-site:4.7.plan-approval): Final sign-off on the rendered HTML plan before Phase 5 starts replacing the scaffold with real pages, components, and design tokens.
>
> Trigger: Phase 4.3 rendered docs/create-site-plan.html; Phase 4.4 opened it in the browser.
> Why we ask: Phase 5 rewrites the entire scaffold (theme.css, Layout, Home page, components, routes) — undoing that touches every commit in the implementation phase.
> Cancel leaves: Nothing destructive — the scaffold itself can be deleted with the project directory; no Dataverse / deploy fired.
Use AskUserQuestion:
| Question | Header | Options |
|----------|--------|---------|
| Does this plan look good? | Plan | Approve and start building (Recommended), I'd like to make changes |
Output: Approved implementation plan, with an HTML copy committed alongside the project for the user to reference during and after implementation.
Goal: Build all pages, components, and design elements with the chosen aesthetic applied from the start
> Prerequisite: The dev server MUST already be running and verified via Playwright (completed in Phase 2). If it is not, go back and complete Phase 2.
>
> Design reference: Read ${PLUGIN_ROOT}/skills/create-site/references/design-aesthetics.md and apply its principles throughout this phase. All pages and components should be built with the chosen typography, color palette, motion, and backgrounds from the start — do NOT build with neutral styling first and redesign later.
Actions:
Before writing any code, use TaskCreate to create a todo for every piece of work. This gives the user full visibility into what will be built:
/contact)", "Create Dashboard page (/dashboard)"Each todo should have a clear subject, activeForm, and description that includes the file path and what the page/component does. Then work through the todos in order, marking each in_progress → completed.
The scaffold is a temporary loading screen — it must be completely replaced during this phase. Do NOT build on top of it or try to modify the loading animation into a real page. Start fresh with the user's chosen design.
> Narrate progress in the loader: Before each of the steps below, update <PROJECT_ROOT>/public/scaffold-status.json so the user — who may still be watching the Home page loader — sees what's actually happening instead of the hardcoded placeholder cycle. Use a short present-participle message (e.g., "Creating Navbar component", "Creating Contact page"). Include any useful grouping context inline in the message itself. The loader picks up changes within ~1.5 seconds. Updates become no-ops once step 4 replaces the Home page.
theme.css (or styles.css for Angular) from scratch with the chosen color palette as CSS custom properties, Google Fonts, motion/animation utilities, and background treatments. The scaffold's loading screen CSS is discarded entirely. Commit after this step. *Before starting, set the loader status to { "message": "Applying design tokens" }.*{ "message": "Rewriting Layout" }.*{ "message": "Creating <Component> component" }.*document.title on mount to reflect the current page (e.g., "Contact — Contoso Portal"). Use the framework's idiomatic lifecycle hook: useEffect (React), onMounted (Vue), ngOnInit (Angular), or a <title> tag in the frontmatter (Astro). Format: "<Page Name> — <Site Name>", with the home page using just "<Site Name>". *For each page, set status to { "message": "Creating <Page> page" } before writing the file. The loader disappears when the Home page itself is replaced — no further status updates are needed after that.*/ and /about — add all requested routes)index.html (or Layout.astro for Astro) to load the chosen Google Fonts instead of the scaffold's DM Sans + OutfitAI_SUMMARY_PLACEMENTS from Phase 3 is non-empty. For each placement, insert a single comment marker in the target page source at the intended insertion point. No visible placeholder UI, no stub components, no extra routes — just a grep-able anchor that /add-ai-webapi will later find and replace. Syntax depends on the framework:| Framework | Marker syntax |
|-----------|--------------|
| React (JSX) | {/* POWERPAGES:AI-SLOT kind=<pick> */} inside the component's return JSX |
| Vue (SFC template) | <!-- POWERPAGES:AI-SLOT kind=<pick> --> inside <template> |
| Angular (HTML template) | <!-- POWERPAGES:AI-SLOT kind=<pick> --> inside the component template |
| Astro | <!-- POWERPAGES:AI-SLOT kind=<pick> --> inside the component's HTML |
Where <pick> is one of search-summary, data-summarization — verbatim from the placement record.
Placement within the page:
data-summarization (record detail): directly after the page heading, above the detail content — this is where a Copilot-style summary card naturally reads in the reading order.data-summarization (list page): directly above the list / table, below the page heading and any filter bar.search-summary: directly above the search-results list, below the search input — the summary paragraph reads before the keyword hits.One marker per placement, exactly as defined in the marker field of the AI_SUMMARY_PLACEMENTS record. Do NOT add stub components (<CopilotSummaryCard />, etc.), CSS classes, or empty <aside> elements — the slot is just a comment. The site must ship as if AI is not a consideration; the follow-up skill does the real work.
Important: Build real, functional UI with distinctive design applied — not placeholder "coming soon" pages, and not generic unstyled markup. Every page and component should reflect the chosen aesthetic from the moment it's created. The scaffold loading screen should be completely gone after this phase — no trace of the Power Pages branded animation should remain.
Use high-quality photos from Unsplash wherever the site needs visual content. Do NOT use placeholder services (e.g., placeholder.com, placehold.co), broken <img> tags, or leave empty image slots.
How to find images:
WebSearch to search Unsplash for relevant photos (e.g., site:unsplash.com modern office workspace)https://images.unsplash.com/photo-{id}?w={width}&h={height}&fit=cropWhere to use images:
Guidelines:
w=800 for cards, w=1600 for heroes/backgrounds) to avoid slow loadsalt text to every <img> for accessibilityCommit after every individual page and component so breaking changes can be reverted. Each page and each component gets its own commit — do NOT batch multiple pages or components into a single commit.
git add -A
git commit -m "<short description of what was added/changed>"
When to commit:
If something breaks, revert to the last good commit:
git revert HEAD
After each significant change (new page or component), browse the site via Playwright to ensure everything is up to the mark:
mcp__plugin_power-pages_playwright__browser_navigate to reload or navigate to the updated pagemcp__plugin_power-pages_playwright__browser_snapshot to verify the page structure and content are correct — do NOT take screenshotsThe user is previewing in their own browser via the dev server URL shared in Phase 2.7.
Once the scaffold loader is gone, public/scaffold-status.json is just dead weight that would ship with the deployed site. Delete the file from <PROJECT_ROOT>/public/ and commit the removal alongside the final implementation.
> GATE: Do NOT proceed to Phase 6 until ALL customization is complete with design applied. The site must have distinctive typography (Google Fonts — no generic Inter/Roboto/Arial), a cohesive color palette (CSS variables), motion/animations, and all requested pages/features before moving to accessibility verification.
Output: All pages, components, and design elements implemented and verified
Goal: Verify the site meets WCAG 2.2 AA standards using axe-core automated testing and fix any violations
> Prerequisite: All pages and components must be fully implemented (Phase 5 complete). The dev server MUST be running.
Actions:
Install playwright as a dev dependency in the project so the audit script can launch a headless browser. This uses the system-installed browser (Edge/Chrome) — no browser download is needed:
cd "<PROJECT_ROOT>"
npm install --save-dev playwright
Run the audit script via Bash, passing the dev server URL and all site routes:
node "${PLUGIN_ROOT}/skills/create-site/scripts/axe-audit.js" --url <DEV_SERVER_URL> --routes /,/about,/services,/contact --project-root "<PROJECT_ROOT>"
The script launches a headless browser, navigates to each route, injects axe-core from CDN, runs the analysis, and outputs a JSON array of per-route results to stdout. Each result contains violations (with id, impact, description, helpUrl, and affected nodes), passes count, and incomplete count. The script exits with code 1 if any critical or serious violations are found.
Parse the JSON output and record all violations.
For each violation found, identify the source file and apply the fix:
| Violation | Fix |
|-----------|-----|
| Missing alt text on images | Add descriptive alt attributes to <img> tags |
| Insufficient color contrast | Adjust CSS color variables to meet 4.5:1 (normal text) or 3:1 (large text) ratios |
| Missing form labels | Add <label> elements or aria-label attributes |
| Missing landmark regions | Wrap content in <main>, <nav>, <header>, <footer> |
| Skipped heading levels | Correct heading hierarchy (h1 → h2 → h3, no gaps) |
| Missing link text | Add descriptive text or aria-label to links |
| Missing lang attribute | Add lang="en" to the <html> tag |
| Inadequate focus indicators | Add visible outline styles to interactive elements |
After fixing each group of related violations, commit:
git add -A
git commit -m "Fix accessibility: <violation description>"
After all fixes are applied, re-run the audit script (same command as 6.2) to confirm violations are resolved:
critical and serious violations)Present a summary table to the user:
| Page | Route | Violations Found | Violations Fixed | Status |
|------|-------|-----------------|-----------------|--------|
| Home | / | 3 | 3 | Pass |
| About | /about | 1 | 1 | Pass |
| Contact | /contact | 2 | 2 | Pass |
| **Total** | | **6** | **6** | **All passing** |
> GATE: Do NOT proceed to Phase 7 until all pages pass axe-core with zero critical and serious violations. Minor and moderate violations should also be fixed where possible, but are not blocking.
Output: Accessibility-verified site with zero critical/serious axe-core violations
Goal: Ensure the site meets user expectations and all pages work correctly
<!-- gate: create-site:7.review | category=plan | cancel-leaves=nothing -->
> 🚦 Gate (plan · create-site:7.review): Live-site review — last chance to request changes before the deploy prompt. Cancel branch lets the user keep iterating. Fires at step 4 of the action list below.
>
> Trigger: Phase 7 has verified all pages render via Playwright.
> Why we ask: User loses the chance to spot UI issues before deploy; broken pages get pushed.
> Cancel leaves: Nothing — site files stay as-is on disk.
Actions:
browser_navigate + browser_snapshot) to verify all pages load correctly — do NOT take screenshots | Component Type | Count | Details |
|---------------------|-------|---------|
| Pages | 4 | Home (/), About (/about), Services (/services), Contact (/contact) |
| Shared Components | 3 | Navbar, Footer, ContactForm |
| Design Elements | 4 | Playfair Display + Source Sans Pro, 6 CSS variables, fade-in transitions, gradient backgrounds |
| Git Commits | 7 | scaffold + 6 feature commits |
AskUserQuestion:> "The site is ready for review at <dev server URL>. Please check it out in your browser. Would you like any changes?"
browser_snapshotOutput: User-approved site ready for deployment
Goal: Deploy the site and suggest enhancements
> This phase is MANDATORY. Do NOT end the session without asking about deployment.
<!-- gate: create-site:8.deploy | category=plan | cancel-leaves=nothing -->
> 🚦 Gate (plan · create-site:8.deploy): Deploy prompt — invokes /deploy-site on Yes. Skipping leaves the site files on disk for the user to deploy later. Fires at step 2 of the action list below.
>
> Trigger: Phase 8 entry; Phase 7 review approved.
> Why we ask: Auto-deploy picks whatever env PAC CLI happens to be pointing at — wrong-env first deploy is messy to undo.
> Cancel leaves: Nothing — site files stay on disk; no deploy fired.
Actions:
> Reference: ${PLUGIN_ROOT}/references/skill-tracking-reference.md
Follow the skill tracking instructions in the reference to record this skill's usage. Use --skillName "CreateSite". Note: .powerpages-site may not exist for first-time sites — the script exits silently.
AskUserQuestion with options: Deploy now (Recommended), Skip for now:> "Would you like to deploy your site to Power Pages now?"
/deploy-site skill./setup-datamodel — Create Dataverse tables for dynamic content/add-seo — Add meta tags, robots.txt, sitemap.xml, favicon/add-tests — Add unit tests (Vitest) and E2E tests (Playwright)/add-ai-webapi — Add generative-AI summaries (Search Summary and Data Summarization). Recommend first when AI_SUMMARY_PLACEMENTS from Phase 3 is non-empty — the pages already carry POWERPAGES:AI-SLOT comment markers at the intended insertion points, so the follow-up skill's explore step finds them deterministically and the user gets the AI surface they picked during discovery without any page redesign.Output: Deployed (or deployment-ready) site with clear next steps
browser_snapshot (accessibility snapshots) to verify pages; never use browser_take_screenshot as it clutters the user's directory. Give the user the dev server URL for visual preview.Before starting Phase 1, create a task list with all phases using TaskCreate:
| Task subject | activeForm | Description |
|-------------|------------|-------------|
| Discover site requirements | Discovering requirements | Collect site name, framework, purpose, audience, and project location |
| Scaffold and launch dev server | Scaffolding project | Copy template, replace placeholders with defaults, git init, npm install, start dev server, share URL |
| Plan site components | Planning components | Determine pages, components, design direction, and routes while user previews scaffold |
| Approve implementation plan | Getting plan approval | Present implementation plan covering design and pages, get user approval |
| Implement pages and components | Building site | Apply chosen design tokens, create all pages, components, routing, navigation |
| Verify accessibility with axe-core | Verifying accessibility | Run axe-core on every page, fix all critical/serious violations, re-verify until passing |
| Review with user | Reviewing site | Navigate all pages, share URL, get user feedback, apply changes |
| Deploy and wrap up | Deploying site | Ask about deployment, present summary, suggest next steps |
Mark each task in_progress when starting it and completed when done via TaskUpdate. This gives the user visibility into progress and keeps the workflow deterministic.
Every site must meet these standards before completion:
"Create a partner portal for our consultants"
partner-portal in current directoryhttp://localhost:5173#1e3a5f primary, blue-gray palettedocs/create-site-plan.html via render-createsite-plan.jsbrowser_evaluatehttp://localhost:5173, requested minor color adjustment/deploy-siteBegin with Phase 1: Discovery
Take microsoft/create-site from the repository into ~/.claude/skills for personal
use, or into .claude/skills inside a project.
The agent identifies a skill by the name field in its header. Two skills with the
same name cannot sit side by side — one of them will be ignored.
The instructions reference npm.
Without those the skill loads but fails at the first command.