Vibe Coding in Action: How AI Collaboration Built an Award-Winning Sleep Aid Product

A Sleep Aid Product Born from Real-Life Observation

Bilibili creator "知识维金" shared an inspiring story: she independently built an AI sleep music companion app called "生息花园" (Breath Garden) using Vibe Coding, and took third place at the Tencent Music Hackathon.

Vibe Coding is a programming paradigm coined and popularized in 2025 by OpenAI co-founder Andrej Karpathy. The core idea is that developers describe their intent in natural language, let AI generate and iterate on the code, and step back from line-by-line micro-operations — instead taking on the role of directing the vision and approving the output. This paradigm's rise is closely tied to the rapid improvement in LLM coding capabilities. When AI can understand vague functional descriptions and produce runnable code, the syntax barrier of traditional programming drops dramatically, enabling designers, product managers, and even creators with zero coding background to independently ship product prototypes or fully deployable applications. That's exactly what makes this story so valuable: it clearly maps out the complete path a regular creator can take — from fuzzy idea to live deployment — with AI tools.

The product idea came from a genuine pain point: information overload has made late-night sleeplessness a shared struggle among young people today. Existing sleep aid products either lack design sensibility or require users to keep making choices right before bed, pulling them back into the doom-scrolling loop. Her goal was to let AI handle the entire "transition from wired-awake to sleep-ready" process on the user's behalf, minimizing decision fatigue.

Three Core Design Principles

Breath Garden's main flow has six steps: choose a healing scene spirit (the lead character is a sleep companion called "夜度猫" / Night Cat), have a brief interaction with the spirit (logging mood, physical state, and desired atmosphere), auto-match songs, enter an immersive playback ritual, generate a personalized voice record, and collect or share a postcard. The product is connected to the QQ Music library, and users can also input songs manually.

The creator distilled three core design principles. The first is the emotional anchor: letting familiar songs immediately become the user's emotional anchor to maximize a sense of security — especially critical for therapeutic products.

The concept of an Emotional Anchor comes from NLP (Neuro-Linguistic Programming) and behavioral psychology, referring to the way specific stimuli (sounds, scents, images) can rapidly activate associated emotional states. In music psychology, familiar songs quickly produce a sense of safety because the brain processes known stimuli with less cognitive load, making it easier for the default mode network (DMN) to enter a relaxed state. For a sleep aid product, using songs from a real library that the user already has an emotional connection to is more effective at lowering cognitive arousal than generative music — this is the core design logic behind Breath Garden's decision to connect to QQ Music's real catalog rather than relying on AI-generated audio.

The second principle is the A/B side vinyl design: Side A plays a hushed version of the song layered with pink noise and white noise; Side B strips it down to just the white noise and pink noise, helping users smoothly transition into sleep.

This design is grounded in clear auditory science. White noise is a random signal with a uniformly distributed power spectral density across all frequencies — it sounds like TV static or a waterfall, and is commonly used to mask environmental noise. Pink noise has stronger low-frequency energy that decreases as frequency rises, closely mirroring the spectral characteristics of natural sounds like rain and wind, making it warmer and softer to the ear. Multiple neuroscience studies have shown that pink noise can synchronize brain oscillations during slow-wave sleep, supporting deeper rest. Pairing a quiet version of a song with pink/white noise in an A/B side structure is a textbook example of translating auditory science into product design language.

The third principle is the flower collection loop: every night a user spends with the app gets "properly preserved," creating a complete product experience.

On the monetization side, the creator envisions bringing the real music catalog back into the high-frequency scenario of nighttime companionship to drive return visits and organic sharing, which can then naturally support membership tiers, full vinyl experiences, and artist co-branding as revenue modules.

The Methodology: From Fuzzy Idea to Live Deployment

The most valuable part of this story is the creator's systematic retrospective on the entire AI collaboration workflow. She breaks it down into three key modules: idea convergence, aesthetic definition, and engineering execution — with the core principle being "make every step bounded, verifiable, and reversible."

Idea Convergence: From Chaos to PRD

The creator admits her initial ideas were "incredibly scattered." Through repeated conversations with Claude, she gradually distilled them from a jumble of loose concepts into a core product vision, eventually settling on four healing pathways and selecting the Night Cat sleep scenario as the primary one. She then expanded a single-sentence positioning statement into a complete PRD that clearly defined the core pain point the product was solving.

A PRD (Product Requirements Document) is the central document in software development that bridges creative vision and engineering execution. It typically covers product positioning, target users, core feature lists, interaction flows, edge cases, and acceptance criteria. Traditionally written by dedicated product managers after extensive interviews, competitive analysis, and internal alignment, the PRD's role shifts in an AI collaboration context: it serves simultaneously as a blueprint for engineers (human or AI) and as a "boundary contract" that constrains the AI's generative scope and prevents scope creep. The clearer the PRD, the fewer misunderstandings the AI will produce and the higher the first-pass accuracy of generated code.

Aesthetic Definition: Translating Abstract Feelings into Concrete Language

The most transferable technique from this phase is translating vague feelings into specific language that AI can act on. Centering on the core concept of "bedtime companionship," she distilled keywords like "grounded, gentle, nocturnal forest, picture-book texture," while explicitly excluding "cyberpunk tech vibes" from the prompts to avoid misinterpretation. To convey a sense of whitespace on the page, she described it as "single visual focal point, page not filled to the edges." Consistent color and typography were the top priority for visual cohesion. She also recommends having the AI take on the persona of a "designer" when writing prompts, then iterating repeatedly.

Engineering Execution: Incremental Verification, Always Ready to Roll Back

This is where the creator hit a classic pitfall: she once asked the AI to "just upgrade the whole thing to make the product look smarter and more premium," trusted it completely without checking in along the way, and ended up with the main user flow broken, all her painstakingly tuned visual details gone, and a flood of strange fonts injected throughout.

Learning from this, she switched to pulling back each decision and having the AI modify things block by block, personally confirming the copy and scope of each change before implementation and sign-off. V1 took roughly three days to complete; V2 focused primarily on integrating the API and adding the flower-language postcard sharing feature. The rhythm of "incremental verification with human oversight" is the key to keeping AI collaboration from spiraling out of control.

AI as an "Autonomously Divided Creative Assembly Line"

The creator's tool stack is worth examining: V1 was built with Claude and Trae — Claude handled coordination and complex product strategy, while Trae handled front-end implementation and scaffolding. When her Claude account was suspended mid-project, V2 was completed with Codex; she noted "no significant difference" in aesthetic output. Background music and the Night Cat's voice assets were all generated by MiniMax.

Claude is a large language model series developed by Anthropic, known for its long-context handling and precise instruction-following. It excels at complex product strategy formulation and maintaining consistency across multi-turn conversations, making it one of the most commonly chosen AI assistants for Vibe Coding workflows. Trae is an AI-native IDE (Integrated Development Environment) launched by ByteDance, deeply integrated with code generation and debugging capabilities and specifically optimized for front-end development. The division of labor — Claude for high-level strategy and content generation, Trae for front-end code execution — embodies the classic "brain and hands separated" collaboration model in Vibe Coding workflows. MiniMax is a leading Chinese multimodal AI company whose voice synthesis and music generation capabilities are among the best in the industry.

The creator's core insight is that AI can become an "autonomously divided creative assembly line." This aligns with a consensus forming among Vibe Coding practitioners — AI is no longer just a code completion tool, but a multi-role collaborator capable of contributing from planning and design all the way through to implementation. But she repeatedly emphasizes that the person who decides product direction, makes trade-offs, and signs off on aesthetic choices is always the creator themselves. The entire process demands continuous critical judgment.

Hackathon Takeaways and Honest Reflection

A Hackathon blends "Hack" and "Marathon" — it's a competitive event where participants individually or in teams rapidly build and demo product prototypes within a set timeframe (typically 24 to 72 hours). The Tencent Music Hackathon is a corporate-themed hackathon where participants usually create within a framework of APIs, data, or scenarios provided by the organizer, with scoring that typically balances creative value, technical completeness, and commercial viability. The spread of AI tools is visibly reshaping the hackathon ecosystem: development work that once required a multi-person team can now be completed solo with AI assistance in the same timeframe, which means creative insight and judgment are becoming relatively more important than pure engineering execution.

On the demo pitch, the creator's view is that the most important thing is aligning the product's value proposition, technical completeness, and the organizer's scoring criteria as closely as possible. During the competition she encountered a wealth of impressive ideas: someone built a hardware music interaction product combining a Raspberry Pi with literary themes; another participant addressed the specific daily life needs of a family member; there were also interactive murder mystery games and DIY player concepts. This left her reflecting — every genuine observation about life has its place and value in this era.

She was also honest about the shortcomings: user research lacked quantitative backing, and commercial considerations weren't fully thought through. As it stands, the product "feels more like a personal passion project than something that can genuinely weave itself into people's daily lives." This candid self-critique is exactly the issue many AI creators need to face honestly — lower technical barriers make "building something" easy, but "building something valuable" still demands a deep grasp of real user needs.

What This Means for AI Creators

The greatest value of this case isn't in how clever the product itself is, but in how clearly it demonstrates a replicable path: start from a genuine pain point, use a single-sentence positioning to converge your idea, lock in your aesthetic before building your flow, replace blind trust with incremental sign-offs, and communicate your value clearly in the final pitch.

The creator also shared a simple but effective habit for building up a creative reservoir: jotting down good product decisions and everyday life observations in a notebook. In an era where AI has dramatically lowered the barrier to creating, what may truly be scarce is no longer the ability to execute — but this kind of life-rooted insight and aesthetic judgment.

Key Takeaways