Skip to content
Beyond the Demo: Shipping an AI Guessing Game to Live Streaming (7.9 Rounds/Day)

Beyond the Demo: Shipping an AI Guessing Game to Live Streaming (7.9 Rounds/Day)

Huajiao Technology details taking an AI-powered guessing game from demo to production in their live streaming app, covering mechanics, AI-assisted development, two major reworks (legacy code modification and position calculation bug), observability improvements, and post-launch metrics showing 7.9 daily rounds per room with 77% from repeat plays.

Huajiao Technology
Huajiao Technology
Huajiao Technology

01 | What Is the Gameplay?

The system presents a riddle with the answer hidden. Viewers submit ideas in chat; the host curates and submits questions like "Is it a daily item?" The AI answers "yes," "no," or "unsure" based on the answer and reference facts. Players discuss and narrow down, or submit direct guesses. Each round allows ten attempts. Valid questions and formal guesses consume attempts; invalid input or AI failures do not. After a correct guess or exhaustion, the answer is revealed and the host can start the next riddle.

Host drives pace, audience contributes ideas, AI provides Q&A feedback. Remaining attempts, permissions, and win detection are handled programmatically by fixed rules.

Live streaming guessing game panel showing riddle, Q&A feedback, and remaining attempts
Live streaming guessing game panel showing riddle, Q&A feedback, and remaining attempts

Host and audience must see the same round progress. Late joiners need to catch up to the current riddle. Scene switches, exits, and consecutive rounds must integrate with existing room operations.

Among 11 concurrent party-room game modes, "Everyone Guesses Together" ranked 2nd by launch coverage.

02 | Clarify Scenarios Before Deciding First-Version Scope

The initial demo already had model Q&A, ten-attempt limit, invalid-question handling, history log, and answer reveal. It could demonstrate gameplay but lacked integration with real rooms and the complete host-audience flow. Demos follow a fixed script; live rooms face unpredictable operation orders.

The team decomposed the gameplay, used AI to organize the technical design, listed unconfirmed questions, and cross-referenced front-end and back-end repositories for reusable code. Riddle types and content were confirmed by business and operations. They also verified that the page would not obscure chat and that host question submission was ergonomic.

Starting a game added a layer of preconditions: room, host identity, mic seat, and page might not be ready simultaneously. If the game started immediately on entry, subsequent operations might not connect; if it waited indefinitely, the host wouldn't know what to do.

Therefore, they first completed the direct-launch flow inside the room, then added the waiting-preparation entry mode. When preparation is needed, the page tells the host it's waiting; once conditions are met, the host clicks to start. Q&A rules are shared; entry, start, and exit are validated separately; audience follows current progress.

The first version omitted audience voting and personal leaderboards. Priority was given to wiring host submission, audience viewing, AI feedback, and continuation after AI failure. Each added scenario could reuse the established Q&A rules without juggling multiple new operation sets.

03 | Turning Design into a Deliverable Gameplay

Using an AI coding assistant, the team built the live-room page, backend services, and model integration. When encountering unfamiliar code, they first asked AI to locate existing handling patterns, then implemented and verified after confirming the approach.

Roughly half a day for technical design, another half day for core front-end and back-end code deployed to test environment. Thereafter, business confirmation, integration testing, and fixes continued until official launch.

Implementation separated two concerns: AI handles understanding questions and generating answers; program logic maintains attempt counts, rules, and results. The page receives host actions; the backend ensures everyone sees the same round progress. When AI fails to answer, the system must inform the host so the round can continue.

Delivery process from gameplay design to real launch showing iterative problem fixing
Delivery process from gameplay design to real launch showing iterative problem fixing

QA-provided historical gameplay test cases served as reference for checking existing room operations. For the guessing game, they added checks for host vs. audience operation differences, premature answer exposure, AI failure recovery, and progress restoration on re-entry.

Pre-launch verification had to answer: If round startup fails, can we pinpoint which step blocked? Asking AI to review runtime logs revealed startup involves room layout adjustment, gameplay progress update, and notification sending. A single "initialization failed" log left the failing action unknown.

They added per-step result logging, enabling correlation of issues within one operation, while avoiding writing user input, answers, or raw AI responses into general logs. Fix scope was limited to guessing-game handling; they re-verified failure and duplicate-operation behavior.

After these changes, page and backend ran respective tests and checked for regressions in older game modes. At release, they confirmed the new page and backend binaries were actually updated.

04 | Two Pitfalls Encountered During AI-Assisted Development

Pitfall 1: AI Modified Unrelated Legacy Gameplay While Building the New One

During change review, they discovered AI had also edited files of the legacy barrage gameplay. AI explained it wanted to prevent legacy gameplay's delayed end notification from incorrectly closing a newly started guessing round. The rationale sounded reasonable, but the changes spread to other legacy gameplay termination processes. A single new gameplay development began dragging in a batch of historical logic adjustments.

The team asked AI to explain each modification's relation to the guessing game and judge which were mandatory for this version. Having a reasonable optimization reason does not mean it must be done now. They reverted unrelated changes in the development phase, kept only the necessary guessing-game integrations, re-audited the diff, and verified original game modes still worked.

Pitfall 2: Position Values Changed, Tests Passed, but Phone UI Didn't Move

Another adjustment targeted the gameplay panel's position on mobile. AI changed numeric values, updated tests, and both checks and compilation passed, yet the panel remained too high during manual testing.

Further investigation revealed the first modification adjusted the lower bound of a generic positioning system, but this page used a separate position calculation. The value actually used on mobile was not the one just changed.

Tests for this change only verified the new numeric value appeared in code. Implementation and tests shared the same misunderstanding, so both passed while the visual result missed expectations.

They relocated the truly effective calculation, restored other pages' original positioning, adjusted only the target panel, and verified the generated page file contained the new rule. Verifying here only confirms the new rule entered the page file; final position still requires on-device checking.

In another debugging round, the phone showed no change because it was still loading old page resources. That was a separate caching issue, not to be conflated with the calculation error, and could not be fixed by further tweaking position numbers.

05 | Post-Launch Observations as Foundation for Next Delivery

After launch, they monitored whether rooms tried only once or continued to next riddles. Over a consecutive seven-day window:

Seven-day consecutive multi-round launch data: 7.9 average rounds, 21.2% rooms with 10+ rounds, 77.0% of all rounds from 'play again'
Seven-day consecutive multi-round launch data: 7.9 average rounds, 21.2% rooms with 10+ rounds, 77.0% of all rounds from 'play again'

Rooms that launched at least one round averaged 7.9 rounds per day, with 21.2% launching ten or more rounds; across all launches, 77.0% originated from "play again."

These records show existing rooms continue to start new rounds. Whether riddles are fun and whether hosting is laborious still need qualitative feedback. They have discussed a survey plan with operations but distribution and collection are not yet implemented.

This delivery also produced reusable assets for the next gameplay:

Integration points and templates. Where gameplay pages and backend connect, how different entry modes start and end — all documented. Next time can start from these entry points without re-scanning the entire repository.

Operation rules and test scenarios. Host vs. audience differences, AI failure, duplicate operations, re-entry, room switching, and exit — all have checklists. New gameplays can reuse relevant scenarios and add their own rules.

Issue logs and debugging experience. Which steps to check on startup failure, what to inspect on page update failure, how to confirm mobile uses the new version — concrete troubleshooting records exist. The missed checks from the two reworks can be carried forward.

Experimentation on the next gameplay continues.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

live streamingAI-assisted DevelopmentobservabilityAI Engineeringproduction deploymentguessing gamepost-launch metricsrework analysis
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.