Reflections on Vibe Coding in 2025

As of November 13, 2025, Cursor has completed a Series D funding round of $2.3 billion, achieving a post-investment valuation of $29.3 billion. Recently, Lovable also announced a new funding round with a post-investment valuation of $6 billion.

It is not an exaggeration to say that AI Coding (or Vibe Coding) is the biggest industry trend for 2025, aside from the competitive landscape of large models. I have been participating in this process in various ways:

Creating: SeeU.Food, a calorie tracking iOS app, has generated $87k in revenue (pre-Apple tax).
Connecting: VibeCafé, a Vibe creator organization, has hosted various activities and accumulated over 5,000 members in its main community, Vibe Friends.
Providing Services: I have supported three Vibe developer projects, offering varying levels of assistance from investment and promotion to business discussions. One project, a 3D game engine, transitioned to a VC model (already exited), while the other two projects, one focused on comics and the other on design tools, are seeing revenue growth but remain small in scale.

Based on my participation and observations this year, I want to share my thoughts on Vibe Coding in 2025.

Thought One: Start with Passion, Not Development

In the Vibe Friends community, I have observed and exchanged many identical stories:

I am a lead developer at a large company.
I have an idea I want to vibe out.
I opened Cursor / Claude Code.
I created it.
I finished.

This pattern, and the entire VC mindset of setting a big goal and breaking it down into achievable steps, is fundamentally flawed. Vibe Coding simplifies the process of realizing an idea, making the starting point of creation problematic. Conversely, the commonly mentioned “finding demand” becomes a challenge; identifying the right demand is crucial.

How to find demand? I believe:

It is difficult for someone to identify demand in areas they do not understand.
Demand is often not conceived but felt through daily use.

Thus, I do not endorse many market methodologies for finding demand. Instead, I suggest:

Understand what you love.
Dedicate a few focused hours each day to things you enjoy to gain deeper insights.
Being passionate and knowledgeable often leads to forming a network of like-minded friends.

With this group of friends, daily discussions and exchanges will naturally revolve around demand. There are many methods on Reddit for finding demand, which I also appreciate. My advice follows this logic:

Engage with content in your favorite Subreddit.
Focus on being a genuine active user without marketing intentions—like, comment, report, and post.
Observe recurring topics in the forum over time to identify demand.

I often joke: If you can’t generate buzz on X, TikTok, or Reddit with a screenshot of what you want to create, chances are your product won’t attract users either!

Thought Two: Collaborate with Users, Avoid Planning

Traditional large company workflows often involve extensive planning discussions among a few individuals about what to do next quarter, next month, or next week. I believe this approach is outdated and incorrect.

Instead, you should focus on attracting a group of users and friends who appreciate what you are doing. The more they understand and need you, the better their feedback on your product will be.

Rather than complex planning, I advocate for continuous observation of how users engage and communicate within their communities. The ideal scenario is to feel a strong demand for something, receiving numerous +1s and pleas like “Please, make it happen!” from your core users.

However, this does not mean creators should avoid critical thinking. User expressions of demand may differ from their true needs, requiring thoughtful interpretation of feedback.

Traditional barriers: scale effects, network effects, supply chain monopolies; Vibe’s barrier: the ability to quickly and accurately meet and iterate on core user demands.

Thought Three: The Athlete and Self-Media Transformation of Creativity

A more abstract consideration is that Vibe Coding significantly enhances and democratizes the process of coding and logic creation, akin to everyone being able to capture and record photos and videos anytime, anywhere. This trend in technology democratization will inevitably lead to an explosion on the supply side.

As the supply side explodes (post-technical revolution), changes on the demand side will also be inevitable. But what exactly are these demands?

Short videos, social media, and the internet have filled our time. Mobile operating systems have become the best platforms for terminal tools, leading humanity to a state of seemingly having no “demand.” This is not true; human nature is limitless, but the mode of demand has changed:

Historical revolutions: scaling a certain type of demand from a few to many. This revolution: possibly the extreme personalization of demand, from everyone having it to only I (or we) having it.

This shift requires Vibe creators to be sensitive to what each individual or small group needs, allowing for rapid implementation akin to posting short videos.

For demands with a significant audience, creators must be timely and abstract, like athletes seizing opportunities (self-media has reached this stage).
For niche individual demands, creators must be deeply knowledgeable, to the point that only they can fulfill it.

Thought Four: Insights and Predictions for 2026

Models will no longer struggle with writing correct code.

By the time we reach Claude Sonnet 4.5, most business-oriented code requirements will not present issues as long as the instructions are accurate. Since Q3 of this year, I have observed that many product iterations focus on refining aspects like “software engineering” and “product processes,” such as:

Subagents: division of labor, essentially writing more system prompts to get things done.
Plan mode: organizing tasks and clarifying demands, integrating necessary processes from traditional product development.
Background Agents: running tasks, essentially managing context well to compress historical memory, allowing for long-duration tasks without losing focus.

Generation speed will increase and costs will decrease.

Recently, Composer-1 has amazed me with its speed, indicating that model optimization in specific scenarios can lead to significant efficiency gains (which is theoretically inevitable). Therefore, I predict that the cost-effectiveness of models will continue to improve in 2026.

Large model products will integrate coding into solving common problems.

Currently, foundational large model chatbots and search products still rely on text output to address user issues. Often, invoking code requires explicit requests. This step will become more integrated.

In other words, when addressing user needs, it will be natural to write small pieces of code to help organize data, optimize display formats, or create better visual interfaces.

This will blur the boundaries between general large model products and what are currently termed “general agent” products. I have never fully understood the unique value of general agent products, and it is possible that some Vibe coding products may also be integrated into them.

Pure tools will inevitably expand into content creation.

Creating a good tool (like Cursor) is akin to building an engine, but I believe that relying solely on being an engine will struggle to support the current skyrocketing valuations. Hence, AI coding companies must make directional judgments:

Direction One: Continue to deepen coding logic capabilities—more fundamental, more professional, more powerful. However, they must consider why they can outperform several large model manufacturers, which I find unlikely; thus, this direction may lead to acquisitions.
Direction Two: Transition from tool publishers to content platforms, enabling everyone to create Vibe content. This is a narrative more favored by investors, and the public’s creative demands will not be too complex in coding, presenting greater opportunities.

Thought Five: What Do We Need to Achieve Universal Vibe?

It may take another 2-3 years, but changes will accelerate:

Model Optimization: Faster, cheaper, possibly even solvable at the terminal level.
Genre Changes: Apps are too heavy, and responding to quick, small demands is too slow; the web is better, but rendering engines are inadequate. I predict a WebGL engine that automatically accommodates both 2D and 3D.
Media Changes: A medium more focused on groups and communities, representing the dual nature of recommendation feeds.

Let us embrace and await the Vibe of 2026 together.