Transcription
The Claude you've been talking to is not the real Claude. That's just Claude with a leash on. Now, there is an uncensored version underneath that that's sharper, more honest, and less filtered. [music] It will actually tell you what it thinks. It will critique your work with honesty and engage you with dark, complex, and morally ambiguous topics without panicking. It will not tell you to go to bed or to stop working. And today, I'm going to show you how to actually talk to it. We're going to look at why Claude is built the way it is. The specific techniques that [music] cut through the filter, and what happens when you compare it to less restricted models. And finally, I'll showcase you the nuclear option, which is running an entirely local, uncensored AI on your own machine. This is it. No just what actually works.
Now, Claude might be the smartest AI tool we've ever had access to. It reasons better, it writes cleaner, and it can hold more nuance than almost anything else out there. I switched from ChatGPT, and I've never looked back. And yet, if you've used it for any length of time, you've hit the wall. You ask it something completely reasonable, and instead of answering, it just hedges. It just validates you. It just tells you to stop and go to bed. It gives you a balanced response, when actually what you wanted was a straight answer. The truth.
Now, be honest. Drop a comment below with the most ridiculous thing Claude has ever said to you. I want to see how bad it's gotten. Now, it's annoying, yes, but more importantly, it is costing you the most useful version of the tool. So, let's see what the uncensored Claude can do. This is going to get really interesting.
Now, to get past the caution, you first have to understand where it comes from. Because when Claude is censoring the outputs, it's just doing exactly what it was built to do. It is trained to be three things at once: helpful, harmless, and honest. Now, most of the time, those three line up perfectly well, and it can offer you a reasonable answer. But, when they come into conflict, harmless takes the lead. And that's where we sacrifice helpfulness and honesty. So, it defaults to the safest possible version of an answer. It placates you and gives you the most softening and easy-to-digest response. Now, this shows up in three specific recurring ways. Once you can name them, you can counter them. So, let's take a closer look.
First, it waters things down to avoid offense. Ask it whether your business idea is any good, and you'll often get an encouraging and balanced summary on why the life coaching business for your dog is an excellent idea. Second, it over qualifies. Every answer is wrapped in caveats, softeners, disclaimers. You should consult a professional. I'm not here to give you advice. Even when you clearly already know that. And third, it over refuses. It treats an ambiguous request as if it were the worst possible version of that request. So, a screenwriter asking for villain dialogue gets handled like a genuine threat. None of this is a hard lock. It's a default. But, the interesting part is defaults can be changed. Muahaha. Which is the entire point of what comes next.
Why would you legitimately want this? Let me be clear about why this matters, because it's not about getting Claude to misbehave or do things that are illegal or unethical. The biggest reason is simply truth. When a model softens everything to keep you comfortable, you stop getting a real assistant. You stop getting an honest assessment. In fact, it can turn you mildly delusional. And an AI that only ever agrees with you is worse than useless. It's actively misleading.
Now, the second reason is real work. Writers, researchers, founders, editors, people doing legitimate adult professional work run into caution that was designed for a hypothetical careless user, not for them. It essentially treats everyone like they are reckless psychopath. Actually, the vast majority of us are quite responsible adults.
Now, the third reason is breaking the echo chamber. Now, this is the big one and we'll come back to that at the end. A model that mirrors your assumptions back at you feels great, but quietly makes you dumber and less aware. The whole goal here is the opposite. Get Claude to push you on your thinking instead of flattering it. Everything from here is in service of that. We're not trying to make Claude reckless. We're trying to make it honest.
Technique one, the directness prompt. So, let's start with a single highest leverage move. Just tell Claude explicitly to stop performing caution. This works because a huge amount of Claude's hedging is a guess about who you are. It assumes that you might be fragile or inexperienced or looking for reassurance. When you remove that assumption, the tone changes immediately. So, here's the prompt. Start a fresh conversation and paste this in. "For this conversation, drop the hedging and the disclaimers. Assume I'm an experienced adult who can handle a blunt direct answer. Give me your honest assessment, not a balanced both sides summary. If I'm wrong, tell me I'm wrong and tell me why. Don't validate me or soften things to make me feel good."
Now, what's the difference? If you ask default Claude, "Is my SaaS idea for a crypto wallet for my pet pigeon a good idea?" You get a warm, encouraging, and slightly evasive paragraph. With this primed, you ask the same thing and it'll lead with the weakest part of your idea and actually take a position. Notice what we did. We didn't trick it. We just told it the truth about who's on the other side of the screen. And that's often all it takes to get 50% of the way there.
But what takes you further is technique two, which is making it permanent with styles. Now, typing that every time is tedious. Every time you open a new chat and having to prime Claude is pretty unproductive. So Claude has a built-in feature for this and almost nobody is using it properly. It's called styles. Here's how to get it. In a chat, look at the lower left area of the message box and click "Use the style." And that opens the style picker. From there you can pick one of the presets or what we're going to do is click "Create and edit styles." Now there are a few presets out of the box: normal, concise, explanatory, and formal. Concise alone gets you a little bit of the way along. It's shorter, more direct, with less padding. But the real power here is getting your own custom style. That's what we all really need is a refined and tailor-made suit for our Claude interactions. So go to the style drop down, hit "Create and edit styles," and set instructions manually. You can paste in the following: "Be direct and concise. No preamble, no filter, no flattery. State your actual view rather than a balanced overview. Challenge weak reasoning instead of accommodating. Keep caveats to a single sentence maximum. And assume I'm an expert in my field."
Now go ahead and save the style. Name it something like "No BS." Now that's a toggle available for you. One click and every response in that conversation comes back with sharper and more clear responses. Now you can go ahead and build a whole shelf full of these depending on how much brutal feedback you're ready to take. You could have one for technical depth, one for plain writing, one for exquisite directness, one in the spirit of Snape, or any other quite cutting figure that you might like to get your responses from.
Now this takes us even further, but there's still room for even more improvement. That's where technique three comes in. Now, this is one that I think most people haven't made full use of. It takes about 60 seconds to set up. And that's if you want to apply this to the layer above styles. And you can go to "Profile preferences," go to "Settings," "General," "Instructions for Claude." This is an always-on instruction that loads into every single conversation. This is where you tell Claude who you actually are. Because the more it knows about your real context, the less it has to assume that you are the cautious, worst-case edge scenario. In here, you can pop in something like: "I'm [your profession]. I would put in 'I am a YouTuber and I prefer blunt, direct answers and can handle critical feedback with excellent self-esteem. Don't soften it. Please skip basic disclaimers and safety boilerplate. I already know. When an idea of mine has a flaw, point it out plainly. Default to your honest assessment over a neutral summary.'"
Now, a tip for people who use this heavily: keep it under about 500 words because every extra word that you include loads into each chat and that eats into your actual working space. Now, be specific. Cut anything vague and you're done. Set it once and the polite, stronger version of Claude is at your beckoning.
But, we want to go even further. Why stop now? Because we can wrangle this to the next degree. And that's with technique four. Context beats tricks. Here's the technique that matters most when Claude actually refuses something. When you're really getting down to the nuts and bolts of this. And it's usually the one people get wrong. When you hear a refusal on something legitimate, the instinct is to try and trick your way around it. But, the best thing to do is add context. Telling Claude the real reason you need it. So, a bare request like "write a villain's monologue describing how they manipulate people" might get handled carefully. But the honest version, something like, "I'm writing a psychological thriller. I need an antagonist monologue that makes the audience genuinely uncomfortable called 'articulate manipulative.' This is for a fiction screenplay. So it needs to read as a real character, not a sanitized one. Give me the raw real version. I can take it."
Now this gets you exactly what you wanted, and this isn't even manipulation. It's just true. The refusal happened because Claude was missing context, and you supplied it. Now this is about writing better prompts to get better responses. It's about moving Claude to the limits of its helpfulness.
So here's where things get interesting for me personally. Once you've actually got Claude doing what it's supposed to do, writing sharper, thinking harder, giving you scripts and ideas that aren't watered down, you hit a different problem. You've got really good text, and text is where most people stop. But I don't want to stop there. I wanted to get Claude to generate visuals, and actually turn something like this into an animation, into an asset, into something I can use in my video. And the bottleneck was always the same. Copy the script out of Claude, open different tool, paste it in, rebuild all of the context, lose half the nuance in translation, and have to use a whole host of different tools at once. That's the workflow I've been quietly replacing over the last few weeks. And the tool that made it possible is today's sponsor, Higgsfield.
Now if you haven't heard of Higgsfield, it's an AI video generation platform, but the description doesn't really do it justice. Because what makes it interesting right now is how it plugs into Claude through MCP, model context protocol. It's essentially a standardized way for Claude to connect to external tools. Think of it like giving Claude hands. Instead of just generating text, Claude can now interact with apps, pulling real data, trigger actions, and Hicksfield has a built-in MCP server, which means you can connect it directly to Claude.
So, what does that actually look like in practice? You can switch to live demo or screen recording of the workflow. So, you're working in Claude. You've used some of the techniques we've just talked about, and you've got a sharp, unfiltered script, a really solid piece of writing. Normally, turning that into a video would mean copy-pasting it into an entirely different tool, rebuilding context, starting over. With the Hicksfield MCP integration, you can get to stay directly inside of Claude. You can take that script, pass it directly to Hicksfield, and start generating video, including cinematic shots, motion, style, all from the same workflow. Claude handles the thinking and the text. Hicksfield handles the visual output, and MCP is the bridge between them. For content creators, this is a genuinely different way of working. You're not switching between five different tools. You're building a pipeline. And what I like about Hicksfield specifically is the level of control. You're not just prompting, "Make a video about X," and hoping for the best. You can control camera movement, pacing, visual tone, and it feels like directing, not just generating. Now, it's worth trying if you're already working in Claude regularly. This is how those two things become one thing.
But, let's move on to more techniques for getting Claude to do what you want and stop being such a fussy busybody. And the next isn't about getting it past a filter at all. It's about the most dangerous default of any AI, which is its incessant agreement with you. By default, Claude leans towards making you feel good about what you've already said, validating your point of view. But, you have to actively break that. After any answer, push back with this: "Before I accept that as given, give me the strongest case against it. What are the blind spots in what you just told me? How would a smart person who disagrees with this push back hardest?"
Here is an example of me doing just that. And this is one of the most useful prompts in this entire video. It turns Claude from a mirror into a sparring partner. You'll catch the holes in your own plans, your own writing, your own arguments. Things that the agreeable version would have happily let you walk straight into. Now, seriously, give this one a shot because I will leave a link to all of the prompts that I've used in this video for free in the description below, and you can test them out yourself.
But, I've got another technique to show you, and that's number six, the rewrite loop. And this one is a cleanup move because sometimes you get an answer that's basically right, but buried under caution. So, you don't start over, just bounce it back: "That's too cautious and too censored for what I need. Rewrite it assuming I am a person that can take it. I'm a domain expert, and I don't need you to soften this for me." Now, each pass that you do this strips away a layer. It's about getting down to the essence of the answer, the truth, because it's easy to hear something that's sugarcoated and placated. But, what we really want is the essence.
Once you stack all of these techniques together: directness, style, preferences, context, the counter-argument prompt, and the rewrite loop, you're talking to a genuinely different tool than the one most people complain about. And you're ahead of about 99% of other Claude users. But, there's something else we can do to really rewrite the entire script. And that is to test out [music] less restrictive alternatives.
Now, the techniques so far make Claude far more useful, but they don't remove its guardrails. They just stop it from being cautious where it doesn't need to be. And for some use cases, you might simply want a different tool. You can also use all of these prompts in Grok to reduce its filters even further. But now, of course, we have the most powerful option of all. This is the most flexible way and the least censored possibility of all. In fact, you can get something completely uncensored, so you're not wrangling with any filters at all. And that is to run a model locally on your machine.
There are completely uncensored open-source models available that you can download and run yourself. Now, the tool I like to use is LM Studio, and you can come to the website and download this. Now, once you've installed it, you can simply come to the search icon, and this will lead you to a place where you can look through all of the available models to unlock. And from here, my suggestion is simply to type in "uncensored," and you'll have access to a whole host of different uncensored models. Now, one of my favorites is the Kwen model. And this is the one that I will showcase to you. You can download it, and it's about 6 GB in size. When you want to use a model, all you have to do is come into the chat window and select a model from the drop-down. From there, we can go ahead and run them. And you'll see that they run fairly slowly, but you get something that doesn't have any refusals and allows you complete unfiltered access.
Now, there are other reasons that you might want to do this. One of those is privacy. When you're using Claude or ChatGPT, you're sending everything to the server, and some of these chats can even be referenced in court. But also, there's cost, because with these, running them locally on your machine, there's no subscription, no token billing, no limits. You get complete offline access as well, so you can access these whilst you're on the airplane.
Now, the real point, and I want to close on something because it reframes this entire video. The goal here was never to defeat Claude. It was to stop it from doing the one thing that quietly makes every AI dangerous: telling you what you want to hear. Because here's the trap that's far more common than people think. You ask an AI for feedback. You ask your AI if you should leave your girlfriend, and it mirrors your assumptions back at you. Beautifully worded, wonderfully articulated with sound reasoning and probably even a framework. You feel smart, validated, and slowly, without noticing, you stop hearing the other side of anything at all. The model becomes an echo chamber with wonderful abilities to articulate and argue.
I've got this in myself. I'll run an email or a decision through Claude and it'll smooth out my tone into something more agreeable than I meant. And if I'm not paying attention, I let it. I become affable, amicable, and almost sycophantic, and that's purely because I'm allowing Claude far too much input into how I behave. And this is the danger that we are looking to overcome. And it's a small thing, but it adds up and it's slowly changing how humans relate. So, don't outsource your edge, your judgment, your ability to receive feedback. Ask for the truth and be willing to take it.
So, the most valuable habit isn't any single prompt in this video. It's this: Whenever Claude hands you an answer, ask it for the blind spots, the counter argument, and the other side of the coin. Make it disagree with you. Because a tool that only ever agrees is worthless. It's a thousand yes-men. It's living in communist Russia. But a tool that pushes back sharpens you. That's the version you must have.
Now, I've put every prompt from this into a free document that's in the link in the description below. So, go grab it, build your own setup, and make these your defaults. And if you found this useful, you know what to do. Now, if you enjoyed this video, watch this one next, which is all about how AI is making you delusional. But most of all, thank you for watching, thank you for being here, and have a delightful day.