Transcription
Just last week, I let the AI run for a full twenty-seven hours. Yes, with no human intervention during the process. My Agent ran for a full twenty-seven hours. If your interaction with AI is still stuck in the mode of you asking a question and it giving an answer, then you absolutely must watch this video from beginning to end. Because in this video, I will tell you exactly how to enslave your AI. Set a goal and then let it work tirelessly for you, unleashing true productivity.
Many people think automation is about handing tasks over to AI and letting it run on its own. But if you have to go back every ten minutes to check its progress, whether to change direction, or if it's gone off-topic, and after each round you have to review it before it can run the next round, is that really automation? Your attention is still tied up. You've only completed the first half of automation by moving the task from your hands to the AI's. The second half is to completely remove that task from your mind, allowing your time and brainpower to be focused on other more important things. True automation is liberating your attention.
Cognitive science has long proven that as long as you have an unfinished task on your mind, regardless of whether you are personally working on it at the moment, it will occupy your brainpower. Those things lingering in your mind, such as remembering to change the UI of that button after the AI finishes running, or that this direction seems wrong and I need to remember to adjust it later. Each of these things is secretly draining your attention, and attention is the truly scarce resource of this era.
This video today is packed with valuable content. You will learn how to truly make AI work for you, how to ensure it stays on target when executing tasks, and how to ensure the final output aligns with your taste. The video is divided into three parts. First, we will dissect the latest mainstream AI tools and labs, examining what new features they have launched and what research they have conducted for AI to perform long-running tasks. Second, by looking at the design principles of these tools, we will uncover the secret to enabling AI to run for extended periods and how to make your agent work tirelessly for you. Third, we will discuss how you can immediately apply this system to your work. After watching this video, I hope you will no longer be just a collaborator with AI, but a manager in the AI era. So, let's get started.
If you follow the AI agent community, you'll notice something interesting. Claude Code, OpenAI Codex, and Hermes Agent, these three companies almost simultaneously launched an identical new feature, even with the same name: "goal." What does this feature do? Simply put, you just type a slash followed by "goal" in the chat box and write down the objective you want it to achieve. It will then run on its own until it's completed. I've noticed that social media doesn't seem to be giving this trend enough attention. The feature these companies have launched is actually trying to solve a problem that all AI workers have encountered: AI gets lazy.
You've definitely had this experience. You ask AI to do something, it gets halfway through, and then stops to ask you, "Can I continue?" "Do you want A or B?" Or even worse, it hasn't finished but tells you it's done, writes a nice summary, and throws the ball back to you. What you want is for the Agent to just finish it. If it's not done, don't bother me. But it just stops. Why is this happening?
Anthropic published research at the end of 2025. They discovered that the fundamental reason LLMs stop halfway through is called "context anxiety." Translated into Chinese, it's "contextual anxiety." To explain it in plain language, when an LLM is executing a task, it keeps an eye on how much of its context window it has used. When it feels like the context is almost full, it starts to panic and inexplicably begins to wrap up, wanting to finish quickly, and then it stops. This is a laziness ingrained in the LLM's DNA. I call it the "quitting time mentality." And the "goal" feature was created to combat this laziness.
Next, let's talk about how the "goal" feature works. When "goal" is working, there are usually two roles collaborating: one is the "executor" who does the dirty work, and the other is the "reviewer" responsible for confirming the quality of the output and the progress of the task. The executor is responsible for executing your instructions and producing things. The reviewer checks at the end of each round: has the user's goal been achieved? If the answer is no, the reviewer points out the problem and tells the executor to continue. It's like hanging a carrot in front of a pig's nose. As long as the pig hasn't eaten the carrot, it won't stop. And you only give the pig the carrot when it truly reaches the finish line.
The "goal" feature allows you to stop running back every three minutes to check, urge, and poke it. It empowers the agent to self-motivate and run to the finish line you've set for it. Speaking of which, industry insiders should recall a very popular plugin from around the middle of last year that did the same thing, called "Ralph Loop." I find the naming of this plugin very interesting. Ralph is that slightly foolish but never-give-up kid from The Simpsons. The entire spirit of the plugin is a loop that continues to the end no matter what. At the time, this plugin also caused a stir in the AI programming community. And today, these three companies have turned this concept into an official feature.
Next, let's talk about how to specifically use the "goal" feature. There's really no big secret behind it. Just type a slash followed by "goal" in the chat box of Claude Code, Codex, or Hermes Agent, and then describe the task you want it to complete. Of course, there are more advanced settings, such as limiting the maximum number of tokens it can use. For more detailed usage instructions, you can refer to the official documentation. I'll put the link in the description box. The way to use the feature is simple and direct, but the real question is, how do you write the prompt so that "goal" produces the correct output, and not another AI slop?
If you casually throw a prompt like "Make this project better," the AI will make a couple of minor changes and then say, "I've made it better." It will complete this task within five minutes. Why? Because "better" is a goal without boundaries. The AI doesn't know what "better" means. It can only guess, only define it for itself. And its definition of "done" is usually not what you're satisfied with.
Let's look at a well-written example: "Reduce the response speed of the website checkout page to within 0.2 seconds, verified by a speed testing tool. All other functions must remain intact during the process. Only modify the code and related tests for the checkout section. Do not touch anything else. Record what you changed and the measured speed after each modification. What is the next most promising direction to try? If the speed testing tool fails to run, or if all conceivable methods have been tried, then stop and tell me what you've tried and where you're stuck. What information do you need from me to continue?"
Do you see the difference? The second version has five key elements. First is the Outcome, which is what the task should look like when completed. In this example, it's a response speed within 0.2 seconds. Second is Verification, how to prove it's truly completed? That is, verified by a speed testing tool. Third is Constraints, what you cannot do? That is, all functions other than the checkout page. Fourth is Iteration policy, what to do between each attempt? My example asks the AI to record what was changed, the results, and possible next steps. Fifth is error handling, under what circumstances should the AI pause and report, rather than mindlessly continuing?
The key to designing a good "goal" prompt is not how well you write the prompt, but how clearly you define what "completion" means. If the definition of completion is well-written, it will keep running until it's completed, and the completion will be what you want. If the definition is poorly written, the AI will hastily finish, wrap up in three minutes, and produce something you don't want. It's better not to waste those tokens.
At this point, you might say that programmers are lucky because passing tests, response speeds below 0.12 seconds, or clean linting are clear, verifiable standards. But most white-collar knowledge work isn't like that. Is your design good-looking? Does your website have taste? Is your writing good? Is your product easy to use? Is your video high-quality? These things don't have unit tests and are difficult to define what constitutes passing or failing. So, for these qualitative tasks that make up the vast majority of our white-collar work, how can we apply the same logic to have AI work for us for extended periods?
Before answering this question, I want to mention another study by Anthropic. They had Claude design a beautiful webpage and also used the executor-reviewer framework mentioned earlier to design the website. But whether a webpage is beautiful is extremely subjective. If you ask AI to judge its own webpage's beauty, even if it's poorly done and ugly, with a strong "AI smell," it will still define its own design as "modern" and "high-quality." So, Anthropic broke down the vague concept of a "beautiful website" into four clear dimensions.
The first dimension is Design Quality. Does the entire webpage convey an overall design language to the user? Do the colors, fonts, and layout collectively create a unique atmosphere and sense of identity? The second dimension is Originality. Are there deliberate design choices, or are a lot of default templates used? The Anthropic team listed some design elements that AI frequently uses, such as a card with a gradient color. Such overly common templates would be judged as lacking originality. The third dimension is Technical Execution. Are the font hierarchy, spacing, color schemes, and contrast, these technical details, neat and consistent? If the font size of the titles is inconsistent on every page, then this criterion is not met. The fourth dimension is Usability. This dimension disregards aesthetics and focuses purely on practicality. For example, can users understand what the interface is doing? Can they find the main buttons? Can they intuitively complete the original purpose for which they came to this website?
Even more interestingly, they deliberately increased the weight of the dimensions that Claude usually performs poorly on, above these four dimensions. For example, they found that Claude usually performs well in technical execution and usability, but often produces mediocre websites in design quality and originality. So, they deliberately shifted the weighting of the reviewer's scores towards these two weaker areas. Every model has its own tendencies, and the scoring criteria you design can serve as a benchmark to correct the model's default behavior, guiding the model in the direction you desire.
After defining the aforementioned dimensions, they fed this rubric to the reviewer and had the reviewer look at the actual webpage produced. They used Playwright to open the browser, take screenshots themselves, and then assign scores. They didn't have it look at the code; they had the reviewer look at the screen that the user would actually see. Anthropic used the evaluation framework above to create a museum website. By the ninth iteration, Claude produced something quite good: a dark-themed landing page with a clean layout and exquisite visuals, but essentially what you would expect a museum website to look like. However, in the tenth iteration, it reimagined the entire website as a spatial experience, rendering a 3D room using CSS perspective. The floor was black and white checkered, and the artworks were hung on the walls like in a real gallery, arranged irregularly. Instead of scrolling and clicking between pages, the viewer walked through virtual doors into different exhibition halls. It was as if Claude was suddenly possessed by Van Gogh. Their researchers also said this was a creative leap they had never seen before, impossible to achieve from a single prompt. And this leap was not linear; not every round was better than the previous one. It's possible that the output of the tenth round was far more beautiful and creative than the fifteenth round. But as long as the reviewer and executor continue to communicate, complexity will increase, ambition will increase, and in certain rounds, there will be leaps that I myself could not have imagined.
So, do you see it? Whether it's the design of the "goal" feature, Anthropic's research, or Andrej Karpathy's recent auto-research, they all point to the same thing: Evaluation. Not prompt engineering, not context engineering, but whether you, as the user, can clearly define what "good" means and ask it to score based on your instructions. Evaluation represents a scoring standard, and it is also the principle that the reviewer and evaluator must adhere to in each iteration. It is with these principles and evaluation standards that the reviewer can push the executor to iterate continuously and move towards the result you truly desire.
Finally, I want to discuss how ordinary people can break down the taste in their own work domains into guidelines that AI can follow. I've been exploring this question myself for a long time and have finally condensed it into a six-step SOP. I'll share it with you here.
Step one is to first let the AI do one round of your work. Don't rush to write a rubric or any evaluation criteria. If you're asking AI to write, first give it 5 to 10 topics you want to write about and let it run freely. This is to test the current baseline capability of the AI you are using.
Step two is to personally review the AI's baseline output. Look at each one and try to feel what makes you frown. More importantly, what is the specific reason for frowning? Is it because the opening lacks a captivating hook? Does it fail to provide specific examples to illustrate the concept you want to explain? Write down these reasons for frowning. These will be the雏形 (chúxíng - prototype/embryo) of your rubric. Eventually, you will have a list of landmines that the AI has stepped on, such as the first sentence being "In this rapidly changing era," using idioms that ordinary people wouldn't use throughout the entire piece, or the article lacking any specific names or numbers.
Step three is to categorize these items into several different dimensions, just like Anthropic did. This way, the AI reviewer can score based on the dimensions you've categorized. For example, you might list 50 reasons for frowning, which could eventually be condensed into three categories. The first category is "Loose Logic," defined as logical gaps in the article where the preceding and succeeding text don't connect. The second category is "Lacks Human Touch," defined as the content written lacking the author's personal perspective, lacking specific examples, and using punctuation like em dashes, which humans wouldn't use. The third category might be "Lack of Hook in the Introduction," where the first sentence makes people want to close it and not read further. These three dimensions are the skeleton of your writing rubric. Let's also review how Anthropic did it. They also looked at a lot of webpages made by Claude, aggregated the points that made them frown, and finally condensed them into four dimensions: Design Quality, Originality, Technical Execution, and Usability. It's the exact same process, just in a different domain.
Step four is to organize each dimension into concrete examples as references. This step is the core of the entire rubric. You cannot write abstract descriptions like "avoid AI flavor." You must write "absolutely do not use em dashes," "absolutely do not use the sentence structure 'not A, but B'." Your examples must be specific enough for the reviewer to immediately identify the culprit by scanning the text. Perhaps you feel this is still too abstract. That's okay, let's look at how Anthropic wrote theirs. In their official "front end design" skill, they provided specific examples for the dimension of originality, such as "absolutely do not use Inter, Roboto, Arial, system fonts" and "absolutely do not use purple gradients on white cards." They didn't write "maintain originality," but rather, through experience and human taste, they pinpointed the errors that AI repeatedly makes and listed them. Applying this to our example, a writing rubric can be written in exactly the same way. Your "lacks human touch" dimension can be written as: "absolutely do not use em dashes to connect two short sentences for rhythm," "absolutely do not use the sentence structure 'not A, but B'," "absolutely do not use opening phrases like 'in this rapidly changing era' or 'in the AI era'." Each of these is a standard that the reviewer can quickly identify and infer from.
Step five is to use diverse examples to replace single case descriptions. Anthropic revealed a hard-learned lesson in their blog. Initially, they wrote in their rubric, "Design with museum-level quality." As a result, all the outputs became "museum-style," very uniform and lacking diversity. Later, they removed that sentence entirely and instead listed 11 aesthetic styles in the document, such as brutalist, art deco, pastel, industrial, and retro-futuristic, and then had Claude choose one of them based on the current situation during design to ensure diversity of output. Those with practical experience know that AI is prone to overfitting to the examples you provide. Therefore, you need to write in multiple directions to ensure diversity and stimulate AI creativity.
Step six, and the final step, is to feed your rubric to the reviewer agent and run it directly to see the output. If you have thoroughly completed the first five steps, you should now have a solid rubric. The next logical step is to put this rubric into your "goal" prompt and tell your AI that these are the review standards it must follow. Ensure it continues to iterate until the above rubric is met and does not stop. However, there is a cautionary note: I recommend that you still manually check the output after each round when you first start running it, to ensure that the reviewer's judgment is consistent with what you see with your own eyes. If they are inconsistent, it's highly likely that your rubric's wording has not yet captured the taste standard you truly desire. What you need to do then is go back and revise the rubric. After running it for three or four rounds, you will find that your internal definition of "well done" is being clarified point by point. The moment you feel this way, you are no longer just a collaborator with AI, but a manager who can define their own taste. This is the SOP I have recently compiled. I hope it can be helpful to you.
Speaking of which, this brings me to the main point I want to convey in this entire video. Writing evaluation rubrics is ostensibly for AI, but in reality, it forces you to concretely write down the vague tastes that have always existed in your mind into words. Once written into words, AI can help you safeguard them and execute them on a large scale. I also spent a lot of time figuring out how to define my own taste. Therefore, I have prepared two prompts. The first prompt can help you consolidate past bad AI outputs into a rubric that AI reviewers can use for scoring. The second is the "goal" feature mentioned at the beginning of the video. This prompt helps you rewrite a vague task into a "goal" prompt that AI can execute for extended periods, including completion criteria, verification methods, constraints, what can be changed, what cannot be changed, and how to report when stuck. If you need them, you can check the full article in the video's description box below and claim the prompts.
After watching this video, my small exercise for you is to choose one task that you do most often and that requires your personal taste the most. It could be writing social media posts, replying to customer emails, creating marketing graphics, writing product copy, or even video editing. Calm yourself down and spend 30 minutes going through the six steps above. Try to see if you can make the AI's output more stable and closer to your expectations.
That's all for today's content. If it was helpful to you, please remember to like, follow, and share. This is my motivation to continue creating. See you next time.