AI Product Managers’ Challenge: Technology Comes Before Aesthetics

Through conversations with many founders, the author notes a consensus: building a product today appears low‑bar, yet delivering genuine value is far harder than during the mobile‑Internet era.

In the past, a product’s success was judged primarily on user experience—smooth interactions and elegant design—essentially a matter of the product manager’s aesthetic insight.

Now, in the AI‑native age, user experience is no longer just an aesthetic issue; it has become a core technical challenge. When users interact via natural language, AI products face a double loss of control: unpredictable user needs and unpredictable value delivery.

How can we deliver better UX and value amid such uncertainty? A fixed set of system prompts cannot precisely match every user’s distinct needs, and it is unreasonable to expect every user to become a prompt‑engineering expert.

Although AI is a pervasive “force,” most people still cannot truly master it. Recent Silicon‑Valley debates about “prompt engineering is dead” point to the next evolution of AI products: the experience bottleneck is fundamentally a technical problem involving both model and product‑engineering layers.

The author proposes two overlooked paths. First, at the AGI Playground conference, he argued that AI product logic follows the principle Input > Output . The architecture consists of an input side, a “magic‑box” model, and an output side. Because the model is an uncontrolled probabilistic system, product design must create determinism within this uncertainty, focusing on deep optimization of both input and output.

On the input side, the author advocates a “wide input” strategy, drawing on Andrej Karpathy’s context engineering and Sean Grove’s normative programming . Karpathy stresses that AI failures often stem from “context failure” – insufficient, ill‑formatted, or inaccurate information – and recommends systematic management of prompts, history, memory, and tools, acting as an “information architect.” Grove argues that many users cannot articulate their true intent; therefore, a structured, iterable specification should serve as the primary “source code” guiding AI.

Both approaches aim to bypass human ambiguity: Karpathy by enriching context so the model can reason autonomously; Grove by forcing humans to clarify goals before AI execution. Relying on users to become context masters or logical experts merely raises the barrier, reproducing the shortcomings of traditional prompt engineering.

The conclusion is that the real breakthrough lies in making AI itself capable of handling the “human‑is‑the‑problem” scenario. AI must proactively discover relevant context, infer true intent, and close the loop between input and output, achieving self‑evolution through multimodal perception and high‑resolution capture of life‑stream data.

Thus, the next step for AI product managers is to deepen their technical understanding of model characteristics and engineering, because without AI‑centric “aesthetic” capabilities, traditional product‑manager intuition will no longer suffice.