Communitygithub.com

Glorya-AI/glorya

Erstelle markengerechte Bildanzeigen, Videoanzeigen, Landingpages und Websites mit dem Glorya-Connector. Verwende ihn, wenn der Nutzer nach einer Anzeige, einer Videoanzeige, einer Landingpage oder einer Website für seine Marke fragt oder Glorya erwähnt.

Was ist glorya?

Glorya ist ein KI-gestützter Marketing-Agent, der die Marke des Nutzers bereits kennt (Logo, Farben, Schriftarten, Produkte und Sprache). Der Agent erstellt auf Wunsch Bildanzeigen, Videoanzeigen, Landingpages und mehrseitige Websites, wobei er eigenständig alle kreativen Entscheidungen wie Format, Größe, Sprache und Zielgruppe trifft – der Nutzer muss lediglich eine kurze Beschreibung geben. Glorya arbeitet asynchron: Nach dem Start durchläuft es einen Job-Loop, der auf Nutzerfeedback wartet, falls Glorya eine Entscheidung benötigt (z. B. zur Konzeptfreigabe). Sobald abgeschlossen, liefert der Agent die Ergebnisse (Bild-URLs, Videolinks oder Vorschau-URLs) und kann auf Wunsch des Nutzers veröffentlicht werden. Für Änderungen oder Verfeinerungen kann der Nutzer jederzeit neue Anweisungen geben. Der Connector ist als MCP-Server implementiert (streamable HTTP, OAuth) und funktioniert mit Claude Desktop, Claude Code und anderen MCP-kompatiblen Clients – der Nutzer muss lediglich einmalig seinen Glorya-Account verbinden. Der Skill beachtet wichtige Regeln: Er verwendet die Standardmarke, fragt nicht nach fehlenden Details, erfindet keine Preise oder Angebote und führt doppelte Aufträge nicht aus. Im Fehlerfall (z. B. Plan erforderlich, kein Guthaben, Marke nicht bereit) gibt er klare Hinweise und leitet den Nutzer zur Lösung.

Funktioniert mit~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/Glorya-AI/skills/tree/HEAD/skills/glorya

In Ihrer bevorzugten KI fragen

Öffnet einen neuen Chat, in dem dieser Agent-Skill bereits geladen ist.

Dokumentation

Was macht glorya?

Glorya is an AI marketing agent. It already knows the user's brand: logo, colors, fonts, products and language. You start the work, follow it and share the results. Glorya finishes on its own unless the user asks to review.

Every tool, input, status, error and limit is in references/tools.md.

When to use

  • The user asks for an image ad, a video ad, a landing page or a website.
  • The user wants to change, publish or stop something Glorya made.
  • The user asks about their recent Glorya images and videos (list_assets), the video styles (list_video_formats) or their brand and products (get_brand).

Connect

The tools come from the Glorya MCP server at https://platform.glorya.ai/mcp (streamable HTTP, OAuth with the user's Glorya account). If they are missing, ask the user to connect it: https://glorya.ai/docs/mcp/quickstart. When they connect, they pick a default brand.

Rules

  • Use the default brand: leave workspace empty. Only when the user names another brand, call list_workspaces with query set to that name and pass the matching id as workspace.
  • Start right away. Pass the user's own words as brief. If they gave no details, pass what they asked for, such as "Create an image ad".
  • Ask nothing before you start: not the format, size, language, audience or style. Glorya picks these.
  • Set language, format or references only when the user gave them. references are public https links.
  • Leave review out; it is false, so Glorya finishes on its own. Set review: true only when the user asks to review or approve steps ("let me review", "show me the concept first"). Never ask the user about it.
  • Never invent products, prices or offers. Never start the same job twice.

The job loop

  1. Call the create tool. It returns job_id and open_url. Tell the user it started and share open_url.
  2. Call get_job with job_id and wait_seconds: 30. While status is running, call it again. It already waits for changes, so don't poll faster.
  3. needs_input: Glorya needs a decision. Show each item in questions as written, with its exact choices and preview_images. Wait for the user. Then call respond with job_id, the question's request_id and answer, and go back to step 2.
    • decision and approve: answer is one exact label from choices.
    • input, revise and clarify: the user's own words work.
    • Don't answer for the user. Don't add questions of your own. If the user writes in another language, translate the question but send the exact label.
  4. completed: share the results (see each type below) and open_url. If progress asks the user something, relay it and send the answer with refine.
  5. failed: tell the user what error or progress says. Offer refine with a changed instruction, or a new job.
  6. cancelled: the job stopped. Finished files stay in Glorya.

Changes: call refine with the same job_id and the user's words as instruction, then run the loop again from step 2. Changes use credits.

Image ad

  • Start with create_image_ad.
  • Glorya plans the ad, writes the copy and makes the images. With review: true it stops to approve the plan (concept and copy) and the finished images.
  • When it's done, assets holds the images. Share each url and open_url. progress holds Glorya's latest message, such as the ad copy.
  • For changes, use refine: "Make the headline shorter."

Video ad

  • Start with create_video_ad.
  • Set format only if the user named a video style. Call list_video_formats and pass the matching id. Otherwise leave it out; Glorya picks one.
  • Glorya writes the script, makes the start frames and renders the final cut. With review: true it stops to approve the format, the script, the start frames and the final cut.
  • Rendering takes minutes. Keep calling get_job, and tell the user now and then that it's still working.
  • When it's done, assets lists the video first, with url and poster_url. Share the video link and open_url.
  • For changes, use refine: "Use a calmer voice." To stop it, use cancel_job.

Landing page

  • Start with create_landing_page.
  • The builder may ask one clarifying question (clarify). Relay it like any other.
  • When it's done, share site.preview_url and open_url. The page is not live yet.
  • For changes, use refine: "Add a pricing section."
  • Publish only when the user says so ("publish it", "make it live"): call publish_site with job_id. slug is optional (lowercase letters, digits and dashes).
  • After publishing, call get_job until site.live_url is set, then share it. If site.publish_status shows a failure, tell the user; publish_site can run again.
  • A build can't be cancelled once it starts.

Website

  • Start with create_website. It builds a site with several pages.
  • Everything else works like the landing page: one possible clarifying question, site.preview_url, refine for changes, publish_site only when the user asks, then site.live_url.

When something fails

Tool errors come with a message, a code and sometimes a link. Pass the message and link on. Don't retry in a loop. The full list is in references/tools.md.

  • plan_required: share checkout_url ($99/month Starter plan). Try again after the user starts it.
  • setting_up: the brand is still setting up. Try again in a few minutes.
  • Wrong brand, workspace_required or workspace_not_found: call list_workspaces with query set to the brand the user named, then call the tool again with that id as workspace.
  • limit_reached: up to 3 jobs run at once per brand and 30 start per hour per person. Wait for one to finish, or stop an image or video ad with cancel_job.
  • no_credits: share billing_url so the user can top up.

Never

  • Launch or change ad campaigns. Glorya can't do that from here. Point the user to https://platform.glorya.ai
  • Publish a page or site without an explicit ask.

Examples

  • "Make an image ad for our summer sale" → create_image_ad with that sentence as brief.
  • "UGC video ad for the new serum, and let me review the script" → list_video_formats, then create_video_ad with the UGC id as format and review: true.
  • "Landing page for my other brand, Lumi Coffee" → list_workspaces with query: "Lumi Coffee", then create_landing_page with that id as workspace.
  • "Make the background darker" → refine on the last job.
  • "Publish it" → publish_site, then share site.live_url.

Verwandte Skills