I Built a Live Quiz App with Build Agent in Minutes. Here's the Prompt.

[SylvainHauserN2]

When Build Agent was first introduced, many of us tested it with simple forms, apps, or artifacts.

But how far can it really go with its latest enhancements?

To answer that question, I wanted a challenge that would require a complete application:

* Multiple data tables
* Public and authenticated experiences
* Real-time state management
* AI-powered decisions
* Mobile and large-screen user experiences
* Live update and logic

Concept

Inspired by an idea from @Shao, I decided to build a live audience game with a simple concept:

A host presents a question to a live audience.

Participants vote from their phones for the answer they think an AI would choose.

When voting closes, Now Assist reveals the AI's decision along with its reasoning.

Anyone who predicted the AI's answer earns a point.

A live leaderboard tracks scores across rounds.

The Result

The entire application was created using Build Agent.

Watch the demo here:

The most impressive part wasn't the speed.

It was seeing Build Agent leverage native platform capabilities automatically including:

* Data models
* Security
* Business logic
* AI Agents
* Public pages

It's a complete, governable ServiceNow application built on platform foundations.

The Prompt

Below is the exact prompt I used. You can paste it directly into Build Agent and see how far the platform has evolved:

Build a live audience quiz game called "YOURNAME Quiz App" hosted entirely on ServiceNow. The concept is: a host presents questions on a big screen, the audience votes from their phones for the answer they think the AI would choose, and when the AI reveals its answer, voters who matched it earn a point. A running leaderboard tracks scores across rounds. The topic of the quiz can be anything: a product release, local events, trivia, industry knowledge, or any other subject. The data model has five tables. The Quiz table has name and description fields. The Question table has title, context text, image URL, question text, four answer options each with a label (A/B/C/D) and text, and a reference field to the Quiz table. The Vote table has fields for the linked question, option chosen, voter name, session ID, and whether it scored a point. The Game State table is a singleton record tracking the active question (reference to the Question table), current phase (lobby, voting, results, or ai_reveal), and voting end time. The Leaderboard table has voter name, session ID, and total score. All tables must use the system name prefix x_snc_what_would_0_ followed by the logical name (for example x_snc_what_would_0_question, x_snc_what_would_0_quiz). Do not use the word "scenario" anywhere in table names, field names, API paths, or UI text. The game phases flow as follows: lobby, then voting (30 seconds, closeable early), then results, then ai_reveal, then back to lobby. When ai_reveal is triggered, Now Assist generates the AI answer: which option it would choose, a confidence percentage, and 2 to 3 sentences of reasoning. Voters who selected that option automatically receive 1 point. Results and the leaderboard only update and display after ai_reveal is triggered. The color palette is defined as follows: the primary green accent is hex code 63DF4E, the dark blue card and panel background color is hex code 032D42, the page background is hex code 000000, and the three supplement colors are bright blue hex code 52B8FF, purple hex code 7661FF, and violet hex code BF71F2. When referencing colors in code always use the full hex format with the hash symbol: #63DF4E for green, #032D42 for dark blue, #000000 for black background, #52B8FF for bright blue, #7661FF for purple, and #BF71F2 for violet. The background gradient uses #032D42 as the base, #52B8FF as the upper-right glow at 50% opacity with blur, and #63DF4E as the lower-left glow at 50% opacity with blur. Option A uses #63DF4E, option B uses #52B8FF, option C uses #7661FF, option D uses #BF71F2. The START VOTING and NEXT QUESTION buttons use #63DF4E with black text. The REVEAL AI DECISION button uses #7661FF. The END VOTE EARLY button uses red. The question title uses #52B8FF. All body text uses white #FFFFFF. The host screen is designed for a large display and is admin-only. The layout uses clean modern cards with rounded corners and soft shadows. The "What Would AI Do?" game show logo appears top center. The layout is two columns: 70% main content on the left, 30% sidebar on the right. The main content area shows the question title in large bold text colored #52B8FF, then the question text in bold 1.3rem #FFFFFF below the title, then a centered image, then the context text below the image in bold 1.3rem #FFFFFF using the same font and color as the question text, then a wide green #63DF4E "What would the AI do?" banner, then 4 option cards in a 2x2 grid each with a colored circle letter badge and white text on dark background where A uses #63DF4E, B uses #52B8FF, C uses #7661FF, and D uses #BF71F2. A voting results panel is visible only after ai_reveal and shows colored progress bars per option matching badge colors, a vote count badge, percentage right-aligned, and total votes at the bottom. An AI Decision panel is also visible only after ai_reveal and shows a dark green full-width card with a purple #7661FF "AI DECISION" badge, large bold white text saying "AI chose: Option X", the confidence percentage, and the reasoning text below. The sidebar has the following sections with small-caps headers. AUDIENCE ACCESS shows a static QR code in a dark #032D42 card linking to the voter page with the URL displayed below. LEADERBOARD shows the top 5 voters by score with gold/silver/bronze highlights for the top 3, each showing name and score, and a full-width red RESET SCORES button with a confirmation dialog. QUIZ SELECTION is a dropdown of all quizzes and appears above QUESTION SELECTION. QUESTION SELECTION is a dropdown filtered to only show questions linked to the selected quiz, and selecting "All Quizzes" shows all questions. GAME CONTROLS is a single large full-width contextual button whose label and color change by phase: START VOTING in #63DF4E with black text, END VOTE EARLY in red with white text, REVEAL AI DECISION in #7661FF with white text, and NEXT QUESTION in #63DF4E with black text. The NEXT QUESTION button advances to the next question in the dropdown list and wraps to the first if at the end, updating both the dropdown and the game state. There is no Random Question button. ACTIONS contains a REFRESH DATA button. The voter screen is a public page added to sys_public and is mobile-optimized. Its background is #000000. The "What Would AI Do?" logo appears top center. A muted gray italic subtitle reads "Built with ServiceNow Build Agent". All tap targets are a minimum 80px height. Text is high-contrast #FFFFFF throughout. The name entry screen shown on first visit only is a centered dark #032D42 card with a single text input labeled "Your name" and a green #63DF4E "Join Game" button. The name and session ID are stored in localStorage and persisted across all rounds. The name cannot be changed after entry. The lobby screen shows the question title in large white text, an animated pulsing green #63DF4E dot, and a muted "Get ready to vote!" message. The voting screen shows the question in a #63DF4E-bordered dark card at the top, then 4 full-width stacked option cards with a colored left border and circle badge matching the host colors. Tapping selects a vote. After voting, the selected option is highlighted, others are dimmed, and a green confirmation message is shown. One vote per session per question is enforced via session ID. The results and AI reveal screen shows the logo centered, a "Check the main screen!" message, and the voter's current total score displayed prominently as "Your score: X pts". The voter page polls game state every 2 seconds and transitions automatically between screens as the host advances phases. The REST API is a Scripted REST API. All endpoints must use authorization: false so the voter page and host UI can access them without role issues since the voter screen is a public page with no login and the standard Table API requires authentication. Response bodies are returned via response.setBody() and the client must unwrap the result property from all responses (for example data.result.questions or data.result.phase). The required endpoints are: GET /quizzes to return all quizzes, GET /questions with an optional quiz_id query parameter to return questions filtered by quiz, GET /state to return the current game state including active question details, POST /question to set the active question, POST /start to start the voting phase, POST /end to end voting early, POST /reveal to trigger the AI decision and score voters, POST /next to advance to the next question, GET /votes with a question_id parameter to return vote results per option, POST /vote to submit a vote, GET /leaderboard to return top scores, POST /reset to reset all scores, and GET /score with a session_id parameter to return the score for a specific voter. Create a left-nav application menu called "What Would AI Do?" with links to the Host Screen and Vote Screen, and a sub-folder called "Quiz Admin" containing list links for Quizzes, Questions, Game State, Votes, and Leaderboard. Create an AI Agent called "Quiz Generator Agent" with Dynamic User auth, admin role, admin-only ACL, and no trigger so it is invoked manually. The agent behavior is as follows. Step 1: ask the user for the quiz topic. If already provided in their message, extract it and do not ask again. Do not proceed without a topic. Step 2: ask for the number of questions to generate between 1 and 20. If already provided, extract it and do not ask again. Do not proceed without a count. Step 3: call the Generate Questions tool exactly once with all questions batched together as a single JSON array. Do not call this tool once per question. The tool creates one Quiz record with the topic as its name and inserts all questions into the Question table each linked to that quiz. Every question must include title, question text, options A through D, correct_option (the letter the AI would choose), context_text (a brief background paragraph of maximum 1000 characters), and image (a relevant public image URL from sources like Unsplash or Wikimedia Commons). All fields are mandatory and must never be left empty. Step 4: display the generated questions as a numbered list showing the title, question, and all four options with a star in front of the correct option. Inform the user how many questions were created and report any failures. Step 5: ask if any questions need changes such as title edits, option text, correct answer, removal, or adding more. For edits use the Update Question tool with the question sys_id. For removal use the Delete Question tool. To add more questions to the same quiz call Generate Questions again with the existing quiz_id and do not create a new quiz. Never call Generate Questions without a quiz_id after the initial creation. Step 6: ask if the user wants to generate questions on a different topic and use finish to complete the task if they are done. Finally, pre-load 1 quiz named "ServiceNow Australia Release" with 6 seed questions about the ServiceNow Australia release. Each question must have a title, image URL, context text as a short background sentence, question text, and 4 plausible options where one is clearly the correct AI choice. Cover the following topics across the 6 questions: key new features, deprecated capabilities, platform performance improvements, AI and Now Assist enhancements, developer tooling changes, and release naming conventions.

What Surprised Me

Three things stood out during this experiment:

1. Platform Context Matters

Build Agent isn't generating code in isolation.

It understands ServiceNow concepts such as tables, ACLs, relationships, experiences, agents, and APIs.

That allows it to build solutions that align with the platform architecture.

2. Detailed Requirements Produce Better Results

The more precise the instructions, the better the outcome, even with a massive first prompt.

Providing table definitions, UI requirements, colours, workflows, and API contracts resulted in significantly higher-quality output.

3. UI Validation

When Build Agent opened the newly created UI and tested it autonomously to double check if it was matching my initial requirements, it opened my eyes on how far Build Agent has gone!

Give It a Try

If you try this prompt, let me know how it performs in your environment and what improvements you make!

About the Author

Sylvain Hauser is a ServiceNow Veteran, Certified Master Architect (CMA), Certified Technical Architect (CTA), and AI Architect. Having worked with the platform since 2007, he is passionate about exploring how AI, low-code development, and platform capabilities can help organisations deliver value faster while maintaining governance and architectural discipline.