Summary of the Key Points
This article discusses how, as AI evolves from being capable of answering questions once to performing complex tasks continuously (such as writing reports or conducting research as an “Agent”), simply optimizing prompts is no longer enough. We need to learn about “Context Engineering”—which involves managing all the information that AI has access to at each step of its work process. The core idea is that, just like human attention, AI’s capacity for processing information is limited. Providing too much information isn’t always beneficial; instead, it’s crucial to present only the most relevant and critical details at each stage to ensure tasks are completed accurately and steadily. These methods can also be applied by ordinary users to address issues like AI forgetting information, making repeated mistakes, or providing off-topic responses.
1. Don’t Just Focus on Prompts; the AI’s “Workbench” is More Important
Do you think creating a prompt is just giving an instruction to AI? In reality, a prompt is more like a task assignment, while the context includes everything else that AI is aware of at the time—what you say, previous conversations, reference materials, information AI has retrieved itself, and intermediate results. For example, when editing an email, the difference between the prompt and the context might be minimal (just a few requirements and the original text). However, when asking AI to conduct industry research, it will first search for information, read papers, and compare data, adding new elements to its “workbench” at each step. In this case, whether the workbench is cluttered with irrelevant or conflicting details significantly affects the outcome.
The next time AI gives an off-topic answer, don’t blame it for being careless; instead, check if the workbench is overloaded with old tasks or unnecessary information—just like how a cluttered desktop makes finding things slower.
2. AI’s Attention Has a “Limited Budget”; More Isn’t Always Better
Large models are often touted as capable of handling millions of words of data, but don’t try to cram everything relevant into them. This can lead to “contextual corruption”—the more information there is, the harder it becomes for AI to find what’s needed accurately. Why? Because AI’s attention is limited, like a budget; adding each new piece of information consumes some of this capacity. For instance, if you ask AI to revise a 3,000-word article, providing the original text, the revision goals, and one reference style is enough. Adding ten old drafts or dozens of previous conversations will dilute the relevant signals, making it harder for AI to focus on the key points.
More dangerously, even though AI’s language output may seem coherent, its ability to recall facts and reason correctly can be impaired. It might seem like it remembers everything, but in reality, its attention has been distracted by the excess information.
3. Provide Just Enough Information; Less Isn’t Better Either
Since AI’s attention is limited, does that mean less information is better? Not necessarily. The principle here is “minimum sufficiency”: missing one piece of information can affect the outcome, but adding ten more won’t improve it significantly. For example, when researching a company, instead of giving AI 50 reports all at once, provide it with a research goal (like analyzing the company’s profit model), the company name, and three key financial reports, and let it search for the rest as needed. This avoids wasting its attention and ensures you get the necessary information.
So-called “universal prompts” (which cover a wide range of situations) are often ineffective; they’re just unnecessary rules that consume resources without providing real benefits. The better approach is to let AI work first, identify its mistakes, and then add specific rules accordingly.
4. Long Conversations Get Chaotic? Summarize and Start Over
Are you used to having long conversations with AI in the same window? If the conversation becomes too lengthy (e.g., after 30 rounds of editing an article), the most effective solution is to ask AI to summarize the current state (e.g., “The goal of the article is XX, the confirmed points are YY, and the remaining tasks are ZZ”), and then start a new conversation with this summary and the latest draft. This helps by “compressing” the history of the discussion, retaining only the relevant information and removing unnecessary details.
5. For Long-Term Projects: Let AI “Take Notes”; Don’t Rely on Chat Records
When using AI for long-term projects (like a month-long market study), don’t expect it to remember every conversation. Create a “project notebook” that includes goals, completed tasks, important decisions, and standards (e.g., the report should be written in plain language without technical jargon). Each time you interact with AI, provide only the relevant parts of this notebook. For example, if you need to write a user needs analysis, give it the “user research data” and the previously identified requirements from the notebook, rather than going through old conversations.
A Simple Checklist for Ordinary Users:
1. For complex tasks, clarify the goal beforehand (e.g., “Write an article explaining how AI processes information”) and don’t overwhelm it with unnecessary materials.
2. Choose a small but high-quality set of reference materials; don’t pre-load everything that AI can find on its own.
3. When AI forgets or repeats, clean up the context by removing irrelevant content or opening a new window for the conversation.
4. Regularly ask AI to summarize the progress of long conversations and consider starting over if needed.
5. For long-term projects, use an external notebook to keep track of key information.
6. Provide typical examples rather than a bunch of unrelated materials.
7. Before adding more information, ask yourself: “Will this really help AI complete the task?” If not, don’t add it.
In summary, the key to using AI effectively is not to give it more data but to provide it with exactly what it needs—just like keeping only the necessary files on your desktop to work efficiently.