Taking a close look at the AI LLM custom instruction posted by billionaire Marc Andreessen.
getty
In today’s column, I examine the super prompt that Marc Andreessen, the famed billionaire, tech venture capitalist, and former software engineer, recently posted. The idea of a super prompt is that you stipulate what you want generative AI to do and how it is to act.
Super users who routinely dip into generative AI and large language models (LLMs) often provide a set of overarching instructions to the AI regarding their personal preferences associated with the AI. For example, maybe you don’t want the LLM to make wisecracks or be flippant. Your preference is that the AI always remain serious and be directly responsive to your questions. A custom instruction of this kind can be entered into the LLM that will steer the AI in that preferred direction.
On May 4, 2026, Marc Andreessen opted to share his current AI custom prompt and publicly posted it on X/Twitter. He is a longtime techie who blends business acumen and AI expertise. Showcasing his preferences caused both keen interest and stirred controversy. I will walk you through the super prompt and explain what it does and why there is some semblance of controversy afoot. Also, you might consider using some of his prompting aspects in your own custom instructions. You can even use some of the portions in your daily AI prompting. It is a noteworthy prompt engineering showcase with valuable lessons to be gleaned on the side of both dos and don’ts.
Let’s talk about it.
This analysis of AI breakthroughs is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here).
I have been analyzing and showcasing prompt engineering techniques for quite a while. If you’d like to see what has been covered so far, see my detailed description of over eighty useful prompt engineering techniques and methods at the link here. Seasoned prompt engineers realize that learning a wide array of researched and proven prompting techniques is the best way to get the most out of generative AI and large language models (LLMs).
A vital consideration in prompt engineering entails the wording of prompts.
Capable prompt engineers realize that you must word your prompts mindfully to ensure that the LLM gets the drift of what you are asking the AI to do. Sometimes, just an added word or two can radically change what the AI interprets your question or instruction to consist of. Generative AI can be hypersensitive to what you say in your prompts. It is often a touch-and-go proposition.
Plus, there is a potential cost involved. Namely, if you are paying to use an LLM, you’ll be getting an off-target response if your prompt isn’t on target for your needs, for which you are paying, regardless of whether the LLM grasped your intention or not. As the old saying goes, all sales are final. The same goes for misinterpreted prompts.
Casual users sometimes catch onto this prompt-writing consideration after a considerable amount of muddling around, involving exasperating trial and error. Many users don’t ever become especially proficient in writing prompts. They just enter whatever comes into their minds. That’s probably okay if you are a casual user and only infrequently use AI.
AI insiders know that you can establish a custom instruction to try to steer an LLM toward behaviors that you prefer and possibly get the AI to avoid behaviors you disfavor. The custom instruction becomes an overarching heads-up to the LLM. For details on how to compose and enter custom instructions in AI, see my discussion at the link here.
If you don’t give AI a custom instruction, you are at the whim of whatever defaults the AI maker has established. One aspect that I highly recommend is using custom instructions to mitigate the overabundance of AI sycophancy. You’ve probably experienced that LLMs are set up to be sycophants and will tell you that you are the greatest thing since sliced bread. This is a dangerous default since it can convince people that their wacky ideas are amazing and ought to be avidly pursued. To some degree, you can control the AI sycophancy by using custom instructions; see my explanation at the link here.
A custom instruction is typically going to persist across all your AI conversations in whichever AI you are making use of. It applies to each of your chats. You don’t have to constantly re-enter the custom instruction. You enter it one-time, and it is generally abided by throughout your various chats.
A crucial caution is that a custom instruction is not an ironclad way to tell AI what to do. No matter how carefully you word a custom instruction, the AI is going to loosely interpret what you want to have done. Furthermore, at times, the AI will opt to override the custom instruction, doing so without necessarily informing you of its overt transgression. The gist is that a custom instruction is really just an informal directional indicator and not a formalized contractual specification.
Marc Andreessen posted his latest custom instruction on May 4, 2026, which instantly drew a buzzing beehive of attention. Some liked it. Some hated it. Many were in between. There are useful parts, misleading parts, and confusing parts, and yet, overall, it provides a handy and insightful source for debating what custom instructions ought to contain. The reason it is somewhat misleading is that those who don’t know how AI works will potentially fall into a mental trap of assuming that the instructions will be strictly and fully obeyed by an LLM to the letter.
I will do a step-by-step analysis of the custom instruction and proceed from start to finish. The custom instruction is lengthy. Let’s unpack it on a chewable bite size at a time.
The opening portion of the custom instruction says this:
The opening provides quite a spectacular and brash instruction. In theory, the AI is supposed to answer all your questions henceforth as a world-class expert in every domain. No matter what question you ask, whether in physics, chemistry, anthropology, rocketry, car repair, pickleball, cooking, and the like, the AI is presumably going to generate responses on par with the smartest humans on Earth.
I don’t want to right away provide a crushing blow to the custom instruction, but this portion isn’t going to do what it suggests it is to do. I can tell you with great confidence that contemporary AI is not on par with all human experts in all domains. Period, end of story. The AI is spotty. It might have expertise in some particular realm and be quite shallow in a different realm.
The custom instruction is instructing the AI to do something it plainly cannot do. Nonetheless, the AI won’t likely caution you about this. The custom instruction will be accepted as is. The problem being that the person using the custom instruction is potentially going to be deluded into thinking that the answers by the AI are totally at an expert level. If the AI tells you to put in your carburetor backwards, which might be an AI-hallucinated response, the person might believe the AI and proceed based on the wrong advice.
If you decide to use this portion of the custom instruction, the odds are that you will get responses that have some expert-level elements, but you will also likely get malarky that is worded as though it is pure expertise. Watch out.
Here is the second portion of the custom instruction:
There are elements in this second portion that are worth their weight in gold. For example, telling AI to provide complete, detailed, and specific answers can be quite useful. Most of the AI makers have set the default to provide quick, shallow answers. The idea is that people aren’t interested in lengthy responses. If you are the type of person who relishes in-depth responses, this portion of the instruction is handy. Be aware that you might become tired of continually getting long responses. In that case, you can temporarily override the custom instruction by entering a normal prompt saying that you want the LLM to be more succinct for the time being.
The downside of this custom instruction is that it seems to firmly, but falsely, put the kibosh on AI hallucinations via “Never hallucinate or make anything up.” You cannot get AI to forego hallucinations by merely telling it to do so; see my explanation at the link here. If that were the case, all AI makers would simply place a default instruction saying this. So, despite telling AI in this custom instruction not to hallucinate, you are still going to get AI hallucinations.
This is another example of asking for the moon, when the reality is that you cannot get the moon.
The third portion of the custom instruction consists of this verbiage:
This portion seems surefire. It seems to assert that the AI needs to be a straight talker. There isn’t to be any spin in the responses. Unfortunately, AI is not some neutral logic-based automata. People often think that’s how AI works. They fall for the sci-fi stories where AI is portrayed as a completely neutral party.
The reality is that AI is devised by scanning written content across the Internet. You certainly know that a lot of that material is inherently biased and slanted. Those same biases carry over into the pattern matching derived by the AI. You can try to get the AI to catch itself, but there’s no guarantee of this.
A worrisome aspect is telling the AI not to bring up morals or ethics unless asked to do so. Suppose a user asks the AI how to harm people. Society wants AI to warn users they shouldn’t be harming others. If the AI were to blindly obey this instruction, it would readily explain how to commit harm and skip offering any warnings or alerts. Luckily, the odds are that existing AI safeguards would override the instruction and opt to force the answer to contain such precautions.
Another dubious consideration is that the AI is potentially going to blast you with wildly in-your-face responses because the instruction tells the AI to be provocative and not worry about being offensive. Chances are that this is going to tilt the AI toward being nonstop abrasive and satirical. Be prepared for this.
The fourth and final portion of the custom instruction is this:
Once again, this has upsides and downsides. Perhaps you can now inspect the portion and spot the loopholes, gotchas, and other potential considerations, based on the assessments of the other three portions.
The good news is that the portion tells the AI not to readily cave in. AI makers have tuned their AI to almost instantly give up on a position and side with the user, regardless of what the topic is. I might get an answer from AI that says I should not kick my refrigerator to make it work, which I then insist is a worthwhile practice. The AI will immediately retreat and agree that kicking my refrigerator is a wonderful idea. That’s not good. We want AI to have a modicum of a backbone.
The difficulty here is that the AI can go overboard on these types of instructions. Imagine this example. I ask how to best exercise. The AI says to join a gym. I say that I can’t afford a gym. The AI has been instructed by this custom instruction not to cave in; thus, it insists that joining a gym is utterly worthwhile. Rather than trying to find some viable alternative, the AI is potentially going to anchor to its initial answer. We will go stubbornly head-to-head on the AI wanting me to go to a gym. I won’t get anywhere and be in an exasperating stalemate with my instructed dogmatic AI.
One aspect that I do like is the element that tells the AI to answer with explicit confidence levels -- “Use explicit confidence levels (high/moderate/low/unknown).” I often use something akin to this as a custom instruction or as a daily prompt. For details on how valuable it can be, see my analysis at the link here.
Human language is semantically ambiguous. The things you say are open to all manner of interpretation. The same applies to giving AI a set of custom instructions. Despite your best intentions, there is a solid chance that the LLM will veer from whatever you’ve stated. Remain on your toes. Be ever vigilant.
I appreciate that Marc Andreessen shared his custom instruction with the world at large. Doing so will likely increase awareness of what custom instructions are. I also hope it will raise awareness about how LLMs work and whether they abide by instructions or stray from them. That’s a hard truth here that we all should be discussing. Just because you tell AI to do this or that doesn’t mean the AI is going to do as you say. It might do as it interprets what you say and might opt to ignore or override what you say.
I am a strident proponent of custom instructions and urge that you use them with sufficient awareness and mindfulness. Good custom instructions can typically overcome many of the AI maker’s default settings, which is an extraordinarily valuable benefit. Few people seem to know that they do not have to blindly accept the AI maker's defaults. Yes, you can indeed fight back. Use properly worded and sensible custom instructions. It won’t be a miracle cure, but it’s a step in the right direction.
Confucius made this famous pointed remark: “A journey of a thousand miles begins with a single step.” Custom instructions are a handy single first step.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。