mcpField
W
wix-headless-kit

wix-headless-kit

github.com/wixDevelopment36286 installsUpdated today
Repository

About wix-headless-kit

wix-headless-kit is a skill for building Wix Headless sites from prebuilt, verified vertical templates. It targets Wix-managed Astro and React projects, leaving brand, layout, and copy to be designed while providing a reference for other stacks.

Key features

  1. 01Prebuilt vertical integrations
  2. 02Typed data layer and React hooks
  3. 03Astro pages and headless components
  4. 04REST seed scripts
  5. 05Setup, deploy, and install scripts

Use cases

  1. 01Scaffolding a new Wix Headless site
  2. 02Adding a shipped vertical to an Astro or React project
  3. 03Using template code as a guide for another stack

MCP support

For the plugin route, it uses Wix MCP tools including ExecuteWixAPI and UploadHeadlessWebsiteFiles to provision, seed, and release a site.

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/wix/skills --skill wix-headless-kit

After installing, type/wix-headless-kitin Claude Code to confirm it loaded.

Bundled files

References and scripts installed along with SKILL.md.

guides/
api-run.md
capabilities.md
existing-site.md
feedback.md
migration.md
reference-mode.md
install/
agents-md.mjs
AGENTS.md
attach.mjs
background.mjs
bootstrap.mjs
check.mjs
context.mjs
deploy.mjs
env-ignore.mjs
install.mjs
lock.mjs
pins.json
proc.mjs
setup.mjs
templates.mjs

Details

From the project README

Wix Headless Kit

Build a Wix Headless site on shipped, verified code instead of authoring the integration. Each vertical ships the integration itself — a typed data layer, hooks, components, pages, and a seed script that are already correct. On a stack that runs it (Wix-managed Astro, any React project) the code is deployed and the agent's job narrows to brand, layout, copy, and wiring. On a stack that can't run it (a static site with no bundler, a server-rendered app in another language) the same code is the reference: its REST twin deploys for the browser side, and the rules it encodes are what the agent ports. Either way the decisions live in the code; don't re-litigate them.

Scope. Tuned for Wix-managed Astro, and each vertical ships one shape of its solution. Use it as-is when the brief doesn't contradict it. When the brief asks for something that shape doesn't express — or once the site exists and the work turns to managing or extending it — that's wix-docs and wix-manage, not a workaround here.

The model

  • Shipped code is the implementation. Every vertical ships in the repository's wix-headless-templates skill (skills/wix-headless-templates/<vertical>/), not in this skill's folder. node <SKILL_ROOT>/install/templates.mjs prints where that skill is: the sibling folder <SKILL_ROOT>/../wix-headless-templates/ when it was installed beside this skill (the cold start installs both; a plugin may not), else a one-time fetch into <SKILL_ROOT>/templates/ (a git clone of the skill's repository, a second); every script below resolves it the same way. Every templates/... path in this document is relative to that printed root — there is no templates/ folder inside this skill when the sibling exists. The folder stays with the project (only the composed project/ scaffolds are left out of its repository), so a later session reads the version the project was built from. Each vertical holds:
    • app/ — the framework-agnostic core (TypeScript): a data layer that returns plain, serializable DTOs (images resolved to https URLs, prices pre-formatted), React hooks, and routing-free headless components. Works in Astro islands, Vite SPAs, and Next.
    • app-astro/ — a thin Astro overlay: SSR pages that fetch via the core and pass DTOs to islands, with owner-editable item-page SEO pre-wired.
    • seed/ — a build-time REST seed script (plain-data plan in, created content out) plus its SEED.md contract.
    • INSTRUCTIONS.md — the vertical's playbook: file map, what you build, hard rules.
    • project/ — the vertical composed into the Wix CLI's blank Astro scaffold, with its lockfile: what wix create copies for a new site, so the first vertical installs without resolving.
  • One auth seam. All shipped code calls Wix through src/wix/sdk.ts: on Wix-managed Astro auth is ambient (no client, no id); on any other React setup the same file runs a manual visitor client off the public client id in src/wix/config.ts. The deploy step configures this — nothing to wire by hand.
  • Data as-is; presentation is yours. The data layer, hooks, and cart chrome are wired as-is — never rewrite their internals, re-route them through API routes, or re-derive a request shape. When the brief needs something they don't express, read the shipped file that owns it and confirm the contract with wix-docs; never infer one from generated SDK types, package files, or node_modules. A normal caller-permitted operation belongs in a new data-layer function. A privileged operation belongs in a validated server endpoint — see templates/shared/CUSTOM_OPERATIONS.md. The presentation doesn't ship: the vertical's INSTRUCTIONS names the surfaces you design and implement yourself on the shipped hooks, with a skeleton carrying each surface's contract (for storefront: the shop and PDP pages with their islands, and home).
  • NEVER work from training data or memory about the Wix APIs. Not a URL, a path, a version, a header, a field name, a filter key, or a body. Every Wix call you make or write — in the frontend, in a seed, in a build-time read of a site — comes from the official Wix skills installed here, the code they deployed first, or, when they do not cover the call, from the official Wix documentation through wix-docs. Read it there first, then write the call. The test, before every request: the exact path and body appear in the output of a file read or a docs search you ran in this session, and you copy them from that output. Anything else is memory: a file whose output was cut short before the call, a source you remember reading earlier, a call built by changing part of one you did find. A guessed call that returns 400 or nothing is not a step toward the answer; it is the failure this rule exists to prevent, trying the next variant is still guessing, and an empty or error reply to a call that failed the test tells you about the call, never about the site. Keep errors visible while a call is unconfirmed: no 2>/dev/null, no | echo. The shipped code is tested against live sites; a body that looks similar is the one that returns nothing, and the API rarely says why.
  • Never mock, fail loudly, purchases via Wix. Live data or an honest empty state; surfaced errors, not swallowed ones; checkout/purchase always through the Wix redirect session.
  • Optional capabilities are deployed from the plan. A vertical can opt into a shared capability without copying sensitive code. For a normal file upload, add a named capabilities.mediaUpload.policies entry to the plan; Fast ships its client helper, Astro endpoint, dependencies, and generated policy module once. Read templates/shared/CUSTOM_OPERATIONS.md before choosing it. The agent wires the helper to the product UI; it never authors or widens the endpoint. For a site-wide search (a header box with suggestions, a /search?q= page over the deployed verticals' products, services, posts and events), add a capabilities.siteSearch entry; Fast ships its data layer, stores, hooks, components and search page once. Enabling it means the Wix Site Search app is installed on the site by the seed step (install: true; the capability's seed/install.mjs, run after the content seed) — the index fills within about half a minute of the install. Playbook: templates/shared/capabilities/site-search/INSTRUCTIONS.md.
The scripts

<SKILL_ROOT>/install/ holds plain Node scripts, no build step, each with a header comment that says what it does and how it is called. They are the deterministic steps of the run: each copies, composes, installs or seeds, and prints one JSON object (or one JSON event per line) saying what it found and did. Read any of them when you want to know what a step does before running it, or when a step's output surprises you; the templates' seed modules are the reference for the Wix calls a vertical needs, in the order that works.

ScriptDoes
templates.mjsFinds the shipped code (the wix-headless-templates skill beside this one), or fetches it once as a sparse clone; prints its folder.
context.mjsReads wix.config.json and .env.local, pulls the env on the CLI route when it is missing, and prints the folder's shape, the deploy site and the content site, the publisher (cli or plugin), whether npm reaches the registry, and whether the upload tool's S3 host answers from this shell (carrier.s3). Changes nothing.
setup.mjsOne call from the folder and the brief to a project ready for the brand layer: scaffolds (or, with --site --app, copies the composed template and writes the config and env), deploys the vertical's code, writes the agent config files, starts the install, and seeds a site it created.
attach.mjssetup.mjs for a site that already exists: does what wix create does after creating one (OAuth app, hosting, env, wix.config.json) against the site id given, then scaffolds, deploys and installs; seeds only with --plan; with --hosting self, puts the given origins on the OAuth app's allow-list instead of Wix hosting.
deploy.mjsComposes a vertical's code layers into the project for the stack (astro, react, lib, static), adds its dependencies, and keeps the lock valid for npm ci (lock.mjs).
install.mjsThe one dependency install: decides none, npm install or npm ci from the lock and what node_modules already holds; writes .install-exit.
bootstrap.mjsCLI route only: checks the Wix CLI and runs the device-code login, one JSON event per line.
check.mjsCompares this installed skill and its templates with the repository they came from; read-only, prints the commands that would update.
agents-md.mjs, env-ignore.mjsWrite a new project's AGENTS.md, CLAUDE.md and .gemini/settings.json; keep .env.local out of git.
proc.mjs, background.mjs, pins.jsonnpm and npx started the same way on every OS; the detached install and seed; the pinned CLI and template versions.
templates/<vertical>/seed/seed-*.mjs, read-site.mjsThe vertical's seed from a plan, and a read of what the site already holds. SEED.md beside them is the plan's contract.

The run

Needed throughout: Node ≥ 24.18, git, the two companion skills installed beside this one, wix-docs and wix-manage, and one of two ways to Wix. node <SKILL_ROOT>/install/context.mjs prints which one this machine has, as publisher:

  • plugin — no Wix CLI session on this machine; the Wix MCP's tools are in your tool list (ExecuteWixAPI, UploadHeadlessWebsiteFiles: a Wix plugin or connector, connected); and npm reaches the registry, so the kit runs here. The MCP carries the user's login. The site is provisioned through ExecuteWixAPI with the recipe wix-manage/references/sites/create-headless-site.md (the provision call returns the metaSiteId and the site's OAuth client appId); step 3's setup takes both as --site and --app and runs no CLI; the seed is an ExecuteWixAPI script written from the vertical's SEED.md (guides/api-run.md section 4); the release is npx astro build, then the build uploaded with UploadHeadlessWebsiteFiles (step 5), attached in one call where the client shows attachments as a local path, or in two calls with curl to the URLs the first returns, from a shell that reaches the internet (the URLs are on Amazon S3; curl -sI -m 25 https://s3.amazonaws.com answers with S3's own headers where they work, x-amz-request-id and usually Server: AmazonS3, and a sandbox limited to a few hosts gets a proxy's refusal with neither). The tool releases the build to the site's Wix hosting, stores the project's source in the site's Code Store repository, and returns the live URL. The plugin route asks for no login. The carrier is settled before the build, from context.mjs's carrier.s3 and from whether the client hands attachments over as files (the tool's attachments description names the clients that do): with neither, the Astro build has no way to the site and the run is guides/api-run.md, the site, the apps and the seed through ExecuteWixAPI and the frontend as a few small static pages dropped in the tool's files form, written against the kit's REST twins and cores as reference (its step 5); the compose is not run, since every file it writes would have to be written out as text.
  • cli — where the Wix MCP is not connected, or a Wix CLI session is already on this machine: a logged-in Wix CLI (npx @wix/cli@latest whoami; npx @wix/cli@latest login is a device-code flow: surface the URL and code to the user, never read tokens into context). node <SKILL_ROOT>/install/bootstrap.mjs checks the CLI and runs the login when there is none; wix create, wix env pull and wix release do what the steps below say.

The cold-start page, https://www.wix.com/skills/headless-cold-start/headless-kit.md, gets a machine with none of this, the skills included, to that point. The four skills this run reads (wix-headless-kit, wix-headless-templates, wix-docs, wix-manage) are on disk after that install (.agents/skills/, or a plugin's folder), or served by the client's skill tool (text to read, not files: the scripts and the templates run only from disk), or online at https://www.wix.com/skills/<skill>/<path>, where <skill>.manifest.json lists every file with its size and a folder's _content returns its files at once; <SKILL_ROOT> below is the kit's folder on disk. In a folder that already holds a wix.config.json, node <SKILL_ROOT>/install/context.mjs first: it runs wix env pull when .env.local is missing (on the CLI route) and prints the folder's shape (folder.shape, the cases of step 3, with the next for each) and the two identities a project has — the deploy site (the config, where the release goes) and the content site (the env, whose app the SDK client runs as and whose dashboard manages the business). They are one site, except on a migration preview (guides/migration.md), where the env names the site being migrated. Every script here reads that context; the site a call targets is never guessed from the config alone. Then fetch the shipped code once: node <SKILL_ROOT>/install/templates.mjs. It prints the folder; the templates/… paths below are relative to that folder, never assumed under <SKILL_ROOT>. node <SKILL_ROOT>/install/check.mjs says whether the skill or its templates have a newer version and prints the update commands; it changes nothing.

Throughout any run, keep the site working at every step so the user can try it as it grows. Steps only the site owner can take never hold a flow back: drafted content is live, and anything that takes money reaches Wix's checkout before payments are connected. The owner's steps go in the closing with their links, as next steps the user can take or ask for.

Throughout any run: if the user asks to send feedback to Wix, complains or gets frustrated, or the run hits friction of any kind: anything that cost more turns than it should have, whether or not it ended in an error (a confusing error, a doc gap, a seed that had to be re-run, a shipped file that did not cover the brief, a playbook line you had to read the source to understand, a call you had to work out by trial, a workaround you had to invent, a slow or flaky step, a platform gate), offer to relay it to Wix per <SKILL_ROOT>/guides/feedback.md. Default to offering rather than waiting to be asked; send only after an explicit yes, never automatically. Step 5 ends with the same self-check.

  1. Resolve the stack. Default is Wix-managed Astro — take it unless the user names another framework or the directory already holds one. Then, by what the shipped code can run there:

    • React (Vite, Next, …) runs everything shipped — data layer, hooks, components: --stack react. The agent's own files may be JS; the shipped files are TypeScript and build untouched inside a JS project — never strip them by hand; if the brief wants no TypeScript anywhere, say the shipped code cannot meet that and ask before going on. A named framework is scaffolded with its own command first (npm create vite@latest, …), then step 3 runs in that folder.
    • Another bundled JS framework (Vue, Svelte, Solid, plain Vite): the data layer and the framework-free stores run (src/wix/ has no React in it), the React hooks and components don't apply: --stack lib. The agent binds the stores and writes its framework's components against the same contracts.
    • No bundler, or another language — a static site (plain HTML/CSS/JS), a server-rendered app (Flask, Laravel, Rails, …): reference mode (its section below, and <SKILL_ROOT>/guides/reference-mode.md). Nothing from app/ deploys; the REST layer deploys for the browser side, and the server side ports it for its reads.

    What Wix hosting takes, and what each stack needs to be released there. wix release uploads the folder named in wix.config.json (site.outputDirectory) and serves it as files — no SPA fallback, no directory index, no rewrites: / and real files resolve, a clean client-side route answers 404 when loaded directly or shared. Server code runs only as a Cloudflare Workers build, declared as outputDirectory: { client, server }.

    • Managed Astro is the stack the Wix CLI scaffolds: the Wix Astro integration (@wix/astro, ambient auth), the hosting adapter (@wix/astro-wix-hosting-adapter, the Workers build), @astrojs/react with React 18 for the shipped components, and in astro.config.mjs integrations: [wix(), react()], adapter: wixHostingAdapter(), output: "server", security: { checkOrigin: false }, image.domains with static.wixstatic.com. An Astro project made without the CLI has none of that; add it before deploying, and the site serves every route. The integration's current major (@wix/astro 3, as in this skill's templates) requires Astro 7: a project on an earlier major moves to 7 first, or is connected as a React host.
    • React and other bundlers release their own build as files. So routes are hash routes, or one emitted HTML file per route linked by its file name — decided before the first route is written; any URL handed to Wix as a return target must be one the host serves. Verify by loading a deep URL directly, not by navigating from /.
    • A framework whose build is a server (Next, Nuxt, Remix, SvelteKit, …) releases only as a static export (files, the rule above applies), or as a Workers build through outputDirectory: { client, server }; otherwise it is self-hosted, with its domain added to the OAuth app's allowed domains before checkout can return to it.
    • Static (no build): outputDirectory points at the folder the pages live in; a route is a page plus a query-string slug (reference mode).
  2. The seed plan. The brief decides what is seeded; a site this run makes always opens with content:

    the briefthe plan
    supplies the content in any form: a CSV, JSON or spreadsheet, a list in the prompt, a PDF price list, a folder of photos and a text file, a link to their current catalog, anything that names the contentthat IS the plan: map it into plan.json per templates/shared/SUPPLIED-CONTENT.md and the vertical's SEED.md ("Supplied content"), every entry, names and prices verbatim, their images and no others
    describes the content without listing it ("a store for hand-poured candles, four of them", "a dozen FAQ questions in three groups")draft a plan from the description per the vertical's SEED.md (read only that for this; save INSTRUCTIONS.md for step 4)
    says nothing about content ("build me a store")on a site made for this run (create, adopt, config-only) draft a plan per the vertical's SEED.md so the site opens with content — demo content the agent decides on, without being asked — and say in the closing message that it is placeholder content and where the owner edits it. On a site that existed before the run: seed nothing; its content is its own

    An existing site has no plan of its own: when the brief names a site by its id, or the folder's config does, the site holds the content already; the frontend reads what is there (step 3's attach path), and only content the brief supplies or describes is added to it.

  3. Set up the project, in its folder — one deterministic call, the same for an empty folder and for a project already on disk; the folder decides what it does, from a few file facts: wix.config.json, its site.outputDirectory, the migration variables in .env.local, package.json, index.html (at the root or in public/), src/pages/. node <SKILL_ROOT>/install/context.mjs prints the shape it reads and the next for it. The brief is the instruction: what it asks to switch on is installed on the site the folder names, without asking again. Ask only when acting would create a second site for a folder that already has one, or when a cleanup seems needed.

    node <SKILL_ROOT>/install/setup.mjs --vertical <vertical>[,<vertical>…] [--plan plan.json] [--business-name "<Brand>"] [--stack <stack>]
    

    Name every vertical the brief needs in this one call (a store with member accounts is storefront,members): the first one's template scaffolds the project; the others deploy in the same call, so the one install covers them all. A vertical added after the install has started costs a second install.

    the folder holdsshapewhat setup doesseeded by setup
    nothing, or loose files (a CSV, a brief)empty → createwix create with the vertical's composed template, here; --business-name names the site. On the plugin route, --site <metaSiteId> --app <appId>: the ids the provision call returned for a new site, or, for a site that exists with no project (a dropped site), its id and the companion app's id from get-or-create (guides/existing-site.md); the composed template is copied in and wix.config.json and .env.local written for that site, no CLIyes, from --plan, on the CLI route. On the plugin route setup takes no --plan and seeds nothing: the plan is seeded through ExecuteWixAPI (guides/api-run.md section 4)
    a frontend, no config (a package.json, or index.html at the root: someone's Astro, Vite, Next, plain HTML)project → adoptinit in place gives it a new, empty site, then deploys; --stack from step 1 is required; make the project what that stack needs on Wix hosting (step 1) before or right afteryes, from --plan
    a config whose .env.local declares an active editor migration (EDITOR_MIGRATION_STATUS=ACTIVE), with or without the blank Astro starter the download carriesmigration → migratethe shipped code into the starter (or the composed template around a bare config), deployed with the migrated site's app as the client, the install starts; ready_for_brand_layer says mode: "migrate", the parent as siteId, the child as deploySiteId (guides/migration.md)no, ever
    a config, no frontendconfig-only → refusesthe site exists and has no frontend yet: on the CLI route attach.mjs (below) takes the site from the config, reuses its hosting, scaffolds and deploys; on the plugin route setup.mjs --site --app with the config's ids, in an empty folder, then the build released by UploadHeadlessWebsiteFiles (guides/existing-site.md). That config is what init leaves behind, and init always creates a site: this site was made for this run and is emptyyes: draft the plan as for create, then attach.mjs --plan
    a config and a frontend (a package.json, or index.html inside the folder site.outputDirectory names)wix-project → refusesiterate: never scaffold, init or reseed. deploy.mjs <vertical…> --stack <stack> adds a solution (the client id comes from .env.local, the config as the fallback), then one install, node <SKILL_ROOT>/install/install.mjs (nothing when node_modules already satisfies the lock); a change is file edits; then releaseno
    a config, a package.json, public/index.html and no src/pages/ (a site published through the drop flow, on its Wix Dev Machine)static-in-astro → refusesthe pages stay plain files in public/: deploy.mjs <vertical…> --stack static --out public puts the REST layer in public/js/wix/, the pages are wired in place (reference mode), and no Astro pages are added; release with build then release, the URL staysno
    a config and the CLI's blank Astro starter with no work of its own (wix-astro-blank, only the blank template's packages: a Dev Machine's never-released site, a bare wix create)blank-starter → refusesattach.mjs --business-name <brand> --vertical <v…> [--plan plan.json] on the config's site: the vertical's composed template takes the starter's place (its lockfile with it), the install and the seed start detached; build the brand layer while they runyes, from --plan: the site was made for this run and is empty, unless read-site.mjs shows content
    a config, index.html at the root, no package.json (a site published through the drop flow and downloaded)published-staticthe config's site, no init: site/ becomes the upload, the REST layer deploys into site/js/wix/; the next says to move the pages, styles and assets in; release keeps the URLno

    A folder no row names is wix-project: the fill-only path — never scaffold, init, reseed or delete; deploy, edit, release. A wrong guess costs a slower run, never the user's work.

    Who decides the seed: where the site came from, never the brief's wording. A site made for this run is empty by construction, so it is always seeded — with the brief's content when it supplies or describes any, otherwise with demo content the agent drafts per the vertical's SEED.md, without being asked, so the site opens with something to see. Made for this run means: setup created it (create, adopt), or the folder held only a wix.config.json and no frontend (config-only — the config is what init leaves behind, and init always creates a site; attach reports siteOrigin: "init"). A site that existed before the run — named by its id in the brief or by --site, a project linked to it, a migration's parent (attach reports siteOrigin: "given") — holds content the run did not make: seed only what the brief supplies or asks to add (attach.mjs --plan plan.json, or the vertical's seed module from the project root: node <templates>/<vertical>/seed/seed-<vertical>.mjs plan.json, <templates> being the folder install/templates.mjs printed), and never invent content for it. "A new storefront for my toy store" describes the business, not content to add: nothing is seeded. Read an existing site first either way (seed/read-site.mjs). Seeds are additive and idempotent by name; nothing on a site is ever deleted or overwritten, and the result's preexisting[] names what was already there.

    • The brief names a site by id → not this call: read <SKILL_ROOT>/guides/existing-site.md and follow it (read the site, then attach.mjs, which does what setup does against the site given and seeds only with --plan, which that guide says to pass only for content the brief supplies or asks to add; self-hosting and a project already on disk are in there too).

    --vertical is required and picks which shipped code deploys AND which seed runs. The ready_for_brand_layer event says mode (create, adopt, migrate, published-static), shape, the stack, publisher (cli or plugin), and the next for that stack, including how it releases.

    setup scaffolds with --skip-git: it composes its own steps and leaves version control to you / the enclosing repo, so it does not create the scaffold's usual git repo + initial commit (which would otherwise become a nested-repo gitlink if the project lands inside a repo).

    The project is created in the current directory — the folder the entry had you work from, which already holds the installed skills — so the project is self-contained and a later session opened in it finds everything. It refuses if the folder already holds a file the scaffold would write. --subfolder creates it in a new folder named after the business instead, for a current folder that must stay as it is; the skills then sit one level above the project.

    It emits one JSON event per line and returns in ~35s: scaffolds the project from the vertical's composed template (wix create copies it: the code and its lockfile arrive with the scaffold), deploys whatever the folder still lacks (patching package.json with every dependency the code imports), then starts two detached background jobs — the dependency install (install/install.mjs: nothing when node_modules already satisfies the lock, npm install --ignore-scripts over a tree that is there, else npm ci --ignore-scripts || npm install --ignore-scripts) and the seed — whose logs and completion markers are in the events. The final ready_for_brand_layer event carries the project dir, siteId, ready-made dashboard links, and both markers. Relay notable events. On an error event, recover just that step via the manual path below, then continue.

    Recovering one step, or adding a solution later: the pieces run on their own from the project root — node <SKILL_ROOT>/install/deploy.mjs <vertical…> --stack <stack> (the client id is read from wix.config.json), ONE node <SKILL_ROOT>/install/install.mjs (never a second npm install concurrently: two npms in one node_modules race and redo each other's work; setup already started one — wait on its marker), the vertical's seed module per its seed/SEED.md. A code change on an existing project is done when it is released (step 5) and the live URL shows it — not when a dev server or a local build shows it. A management change (a recipe against the site) needs no release; the frontend reads it live. A site released through the upload tool keeps its source in its Code Store repository: a later change, in any session, downloads it (guides/existing-site.md), edits, builds and uploads with the downloaded revision as expectedHeadRevisionId. One site keeps one writer path: such a site is not changed through wix release or a plain drop.

  4. Design and build the presentation while the install finishes — in the project dir from the ready_for_brand_layer event, per the vertical's INSTRUCTIONS.md: set the @theme tokens, brand the chrome, and implement the vertical's creative surfaces yourself on the shipped hooks (for storefront: your product card + grid, shop surface, PDP surface, and the home page) — designed to fit the brief, not copied from the reference components. Read the INSTRUCTIONS and the shared floors — templates/shared/DESIGN.md + templates/shared/CONTENT.md — now (not earlier — their contracts matter only from this step on); the hook/DTO contracts are inlined there, so don't open the shipped files themselves. A frontend that already exists gets the minimum. When the run wires pages that were there before it (a dropped mock, an adopted site, a published static site), do the least that makes them work: connect each control the pages show to the shipped code, and add only what a control can't work without. Anything more the vertical offers is added only when the brief asks for it; name it in the closing message. If the brief needs a core operation that shipped code does not cover, read templates/shared/CUSTOM_OPERATIONS.md before writing it. Use one documented path and implement it; do not reverse-engineer SDK internals.

  5. When both background jobs have completed — the install's marker (.install-exit, containing 0) and the seed's (.seed-exit) both exist — verify the seed succeeded (.seed-exit contains 0; seed-result.json has the created counts for your summary — if non-zero, read seed.log and re-run the seed module manually). Those two seed files exist only when setup started the seed (attach runs none: only the install marker is waited on). When you ran seed-store.mjs yourself (adopt, iterate and published-static runs, reference mode), there is no marker to wait for: the process's exit code is the result and its stdout is the JSON — wait on the process (a foreground run, or wait on its pid), not on a file. Then build & release once (managed), as the next of the ready_for_brand_layer event says for the stack and the publisher: Astro on the CLI route → npx @wix/cli@latest build then npx @wix/cli@latest release; Astro on the plugin route → npx astro build from the project root, then UploadHeadlessWebsiteFiles with the site's siteId, in one of two forms (the next carries the zip commands, with Python's zipfile where there is no zip binary):

    • attached, one call: the project zipped without node_modules (.wix/build-metadata.json and dist/ in it), passed as attachments. The build is released and the source in the same zip is stored.
    • two calls, from a shell that reaches the internet (the upload URLs are on Amazon S3: curl -sI -m 25 https://s3.amazonaws.com answers with S3's own headers where they work, x-amz-request-id and usually Server: AmazonS3; a sandbox limited to a few hosts gets a proxy's refusal with neither, and cannot release a build this way): the build zip (zip -qr build.zip dist .wix) and the source zip, the project without node_modules, dist, .wix, .astro, .env*, .agents (zip -qr source.zip . -x 'node_modules/*' 'dist/*' '.wix/*' '.astro/*' '.agents/*' '.env*' 'build.zip'). First { siteId, prepare: true, size }, size being wc -c < build.zip: the reply gives a curl -X PUT ... --data-binary @build.zip "<putUrl>", one for the source zip when the site's repository takes it, the repository's headRevisionId, and the exact second call. Run the curls as given, then { siteId, uploadKey, sourceUploadId, expectedHeadRevisionId }.

    The service deploys, commits the source to the site's Code Store repository, then releases; a refused commit releases nothing. expectedHeadRevisionId is the revision the project started from: the one it was downloaded at, or for a project started in this run the headRevisionId the prepare returned. Left out, the release works only on a repository with nothing in it yet; otherwise it answers EXPECTED_HEAD_REQUIRED and releases nothing. A 409 CODE_STORE_HEAD_CHANGED means the repository moved since: download it again (guides/existing-site.md), carry the change over, build and upload. A 404 on the second call means the upload expired: prepare again. The result carries the live URL; a release replaces static pages dropped on the site before, at the same URL. Otherwise: React or another bundler → the project's own build, then npx @wix/cli@latest release of the build folder named in wix.config.json (deep URLs must answer 200 directly, step 1); static → release alone. If the install failed, run it once more and then build. Don't build+release mid-flow; backend content is fetched at runtime, so a re-release never "refreshes" seeded data. The run is complete only when the site is released — close with the live URL and the dashboard link https://manage.wix.com/dashboard/<siteId> (the siteId of the ready_for_brand_layer event: on a migration preview that is the migrated site's dashboard, the release URL is the preview's, the original site is unchanged, and you complete the migration once they approve (guides/migration.md) — say all three). Tell the user how to put their own domain on the site, per the wix-manage recipe references/domains/domain-search-purchase-and-connect.md (installed beside this skill, else https://www.wix.com/skills/wix-manage/references/domains/domain-search-purchase-and-connect.md): bought through Wix, or one they already own. When the site takes money through hosted checkout (a cart, a paid booking or rental, tickets, plans, donations) and nothing says payments are already set up, say what a visitor meets at checkout until they are — "We can't accept online payments. Contact us for help with your order." — and hand the two links that fix it: Accept payments https://manage.wix.com/dashboard/<siteId>/wix-cashier/payments (connect a payment method; "manual payments" is enough for free and pay-in-person flows) and Upgrade the plan https://www.wix.com/upgrade/website?metaSiteId=<siteId> (online payments need a premium plan). Both are the owner's steps, not a defect in the site. When the run started from the owner's own pages, name what those pages promised that the released site does not do. Copy the live URL verbatim from the release output (the CLI's, or the upload tool's result) — never retype it from memory (a mistyped subdomain hands the user a 404). Before you sign off, run the feedback self-check over the whole session (guides/feedback.md): anything that cost more turns than it should have, including what you recovered from silently, is signal; if anything qualifies, offer to relay it as you deliver the links, and send only after an explicit yes.

Reference mode — a static site, or a server-rendered app in another language

The shipped data layer exists a second time as a REST layer over fetch, for the stacks that cannot run app/: a static site with no bundler, or a server-rendered app in another language (Flask, Laravel, Rails). The REST layer deploys for the browser side; the server side ports its reads. When step 1 resolves to one of these stacks, read <SKILL_ROOT>/guides/reference-mode.md before step 3: it holds the mechanics (the site/ layout and setup's part in it, what runs in the browser versus the server, the OAuth allow-list for a self-hosted origin, pre-rendered output) and how to close such a run. A public site is still better served by managed Astro; say so when you close.

Where the kit does not run (no route to npm, no Node), or when the install, the CLI or the login is blocked where you are, read <SKILL_ROOT>/guides/api-run.md: the same run, step by step, as the Wix API calls the scripts make and the files beside this skill that carry the contracts. On the plugin route the same guide is the reference for the calls that go through ExecuteWixAPI (the provision call, the seed); the build and the release are this skill's step 5.

Verticals

The shortlist. Match the brief against the first column; when it names a Wix product or a feature not here, when two rows could fit, or when the request sounds like something this skill does not ship (meetings, gift cards, loyalty, groups), read <SKILL_ROOT>/guides/capabilities.md: every Wix product, what it covers, which vertical here ships it and what is not shipped.

The user wants…VerticalPlaybook
Online store: products, categories, variants, cart, checkoutstorefronttemplates/storefront/INSTRUCTIONS.md
Appointments/classes: services, time slots, staff, booking, checkoutbookingstemplates/bookings/INSTRUCTIONS.md
Rentals: rooms, vehicles, gear rented by the hour or the day, customer-picked length, checkoutrentalstemplates/rentals/INSTRUCTIONS.md
Blog: post feed, categories/tags, rich-content post pagesblogtemplates/blog/INSTRUCTIONS.md
Structured content collections (directory, recipes, listings) with pages designed per schemacmstemplates/cms/INSTRUCTIONS.md
Any visitor-fillable form: contact/enquiry, signup, application, survey — rendered from the live schemaformstemplates/forms/INSTRUCTIONS.md
Events: listing, event pages, free RSVP, ticket sales via hosted checkouteventstemplates/events/INSTRUCTIONS.md
Member accounts: custom in-app login/sign-up, gated pages, account pagememberstemplates/members/INSTRUCTIONS.md
Portfolio/showcase: collections of projects, project pages with media galleriesportfoliotemplates/portfolio/INSTRUCTIONS.md
Membership/subscription plans: pricing page, plan detail, hosted purchasepricing-planstemplates/pricing-plans/INSTRUCTIONS.md
Restaurant: menu with photos, online ordering, table reservationsrestaurantstemplates/restaurants/INSTRUCTIONS.md
FAQ: questions grouped by category, search, expandable answers, a link per questionfaqtemplates/faq/INSTRUCTIONS.md
Donations: campaign pages, goal progress, one-time and recurring giving via hosted checkoutdonationstemplates/donations/INSTRUCTIONS.md

Verticals compose: a brief that spans several (a restaurant with a blog, a store with member accounts) names them all in the setup call (--vertical restaurants,blog), so one install covers them; each vertical's seed runs with its own plan. On a project already built, node <SKILL_ROOT>/install/deploy.mjs <vertical…> from the project root adds one, then one npm install, then its seed. A request that matches no shipped vertical has no shipped code: say so in one line, then build it from the Wix API reference through wix-docs (search, then the method page), with the same rule as every other call, on the same project and stack, starting from the closest shipped vertical when one exists (guides/capabilities.md says which).

Comments

Markdown supported