Whoever gets the context gets the user

Since the emergence of large language models, all practitioners have essentially been solving two problems:

1. 1. Improve the intelligence level of the model
2. 2. Provide context for the model

Scenario innovation and engineering optimization are essentially about solving the problem of how to provide "context" to the model more elegantly.

• • Prompt Engineering is the context, the “engineering” part is the optimization of the expression of background information;
• • RAG is the context, using text snippets retrieved from your private knowledge base to augment questions;
• • DeepResearch is the context, processing various articles from the Internet into “research needs”;
• • MCP is the context, no explanation is needed, the C in it is;
• • Agent is context, dify, Coze are also contexts…

The prompt word is just a "demand activator". Whether the model can do a good job after being activated depends on what background information you provide to it.

Those "ordinary people" who fail to make good use of AI, on the one hand, do not think of using AI (scenario problem), and more importantly, they do not provide complete information when giving instructions to AI (engineering problem).

Instructions like “write a hit copy for me” are typical “0-added” expressions.

Essentially, all AI-Native products are doing the following:

Fighting hard, fighting forScene CreativityandEngineering Implementation.

To improve the competitiveness in these two aspects, it is enough to do the following two things:

1. 1. Check the top 100 products on the Product Hount list every day. Currently, more than 90% of them are AI applications in various strange scenarios.
2. 2. Check out the open source projects on GItHub Trending every day. Currently, more than 80% of them are engineering implementations of various AI product developments.

The real core is how you can obtain the "context" of users' needs to solve problems through AI in a certain scenario, and provide it to AI in an appropriate way.

The top product of this week’s Product Hount is Remio, an AI Native note-taking tool developed by a former vice president of NetEase.

This note-taking tool has a very "exciting" function: after installing the browser plug-in, it will silently collect all the web pages you open on the browser and synchronize them to the note-taking APP...

Although scary, it is a very good way to record, synchronize and model user context.

A few days ago, mem0AI, a team invested by OpenAI, open-sourced an MCP service called OpenMemory. Its core value is to achieve memory sharing among multiple AI applications (through MCP).

The messages you sent on the Claude client will be summarized and saved. When you go to the Cursor to issue an instruction again, the AI ​​can extract the relevant content from the "memory" stored in Claude and put it in context.

I mentioned this scenario in a live broadcast a long time ago: whoever can help users collect information and automatically complete it in the conversation when they don’t remember it will be able to get users.

OpenMemory does not require the user's "context", but chooses to store it locally.

But for the general public, convenience is more important than privacy . After all, not everyone can or is willing to deploy such a system themselves.

The technical characteristics of large models determine that "context" completion is a must for making good use of AI: more comprehensive and complete input = better quality and more intended output.

As long as this technical feature of the big model remains unchanged, everything you do next will ultimately help users complete the context in a more seamless, smooth, stable and accurate manner.

Once you understand this essence, it becomes very easy to commercialize many creative ideas.