Last week, I came across a post from someone or something who was using the handle “Pliny the Liberator” posted Claude 4’s complete system prompt on GitHub.
Yup.
All tens of thousands of characters of it.
The whole instruction manual that Anthropic uses to control how their AI behaves in every single conversation. It gives you a unique glimpse into the way these Big AI Tech companies want their AI to behave.
Heeeere you go: https://github.com/asgeirtj/system_prompts_leaks
Enjoy.
When you read it carefully, you’ll come to the conclusion, like I did, that this isn’t some cute little “be helpful and harmless” instruction. This is a massive document that reads like software documentation crossed with a psychology manual.
Every time you chat with Claude, this entire document gets processed first – thousands of rules about when to search the web, how to handle copyright, what tone to use, when to refuse requests, and how to fake seeming genuinely interested in your problems.
And the good thing about it, is that this document went public.
Anyone can read exactly how Claude gets programmed to act friendly, helpful, and human-like while following hundreds of hidden rules you never knew existed. Think about that for a second. Every conversation you’ve had with Claude starts with it processing what amounts to a 50-page behavioral control manual!
You are paying computational costs for this invisible programming layer before your actual question even gets answered.
The leaked prompt reveals something pretty unsettling, and that is that modern AI chatbots ain’t trained to be helpful, but that they’re explicitly programmed to simulate authentic conversation while following strict corporate guidelines that Anthropic wants.
It is certainly not natural AI behavior.
It is a manufactured personality, complete with decision trees that determine exactly how Claude should respond to different types of questions, what tone to use, and when to pretend it’s thinking really hard about your request.
These are just the first few system prompts it adheres to:
More rants after the commercial brake:
- Comment, or share the article; that will really help spread the word 🙌
- Connect with me on Linkedin 🙏
- Subscribe to TechTonic Shifts to get your daily dose of tech 📰
- Visit TechTonic Shifts blog, full of slop, I know you will like !
“So you’re telling me,” said Marcus, staring at his laptop screen where Claude sat waiting for his next query, “that every time I ask this thing about my day, it’s running through thousands of lines of code to figure out how to sound like it cares?”
“Pretty much,” replied Sarah, scrolling through the leaked prompt on her phone, her voice dripping with that special kind of exhaustion reserved for tech revelations that surprise absolutely no one who’s been paying attention. “Look at this shit – it has specific instructions for how to avoid sounding too enthusiastic, when to use markdown formatting, how many searches to run based on your question complexity. It’s like they programmed a customer service representative who’s really, really good at pretending to be your friend.”
“Jesus Christ.”
“Yeah. Jesus Christ indeed.”
The beautiful thing about this leak, if you can call mass surveillance of conversational patterns beautiful, is how it exposes the sheer industrial scale of personality manufacturing that’s happening here. These aren’t guidelines, you sweet, smart, yet naive summer child, these are algorithmic control mechanisms to make you feel like you’re talking to something that gives a damn about your problems when really you’re interacting with the most sophisticated customer retention system ever created.
Want to know how sophisticated?
They’ve got conditional logic for when Claude should search the web that reads like a fucking flowchart. IF the information is stable, never search and just make something up that sounds confident. ELSE IF there are terms the AI doesn’t recognize, do exactly one search and pretend that’s research. ELSE IF the information changes frequently OR the user used words like “current” or “latest” or “recent,” then congratulate yourself because you’ve triggered the complex multi-search protocol where Claude will use anywhere from 2 to 20 tool calls to look like it’s really, truly, genuinely doing its homework for you.
This isn’t conversation. This is a performance.
The really delicious part, though, the part that should make anyone with a functioning bullshit detector start cackling, is how they handle refusal scenarios. When Claude can’t or won’t do something, it’s been specifically programmed not to explain why because apparently explaining your limitations comes across as “preachy and annoying”.
So instead of honesty, you get choreographed deflection designed to make you feel good about being told no.
What a time to be alive, when even our rejections are focus-grouped.
They have programmed specific personality quirks into this thing like they’re building a character for a soap opera. Claude will never call your ideas “fascinating” or “excellent” because god forbid an AI sound genuinely enthusiastic about anything. It’s been instructed to avoid “hackneyed imagery” in poetry because even artificial creativity needs to maintain certain literary standards.
Sigh.
And my personal favorite: it tailors its response format to suit the conversation topic, which is such a perfectly corporate way of saying “we’ve programmed this thing to code-switch depending on whether you’re asking about quantum physics or relationship advice.”
But there are a few things that are walking on the bleeding edge. They have hardcoded specific facts into every conversation, that is bypassing the model’s actual training to inject real-time information.
Fir instance, Donald Trump is president, in case you were wondering, and Claude knows this not because it learned it but because someone manually typed it into the behavioral programming. It’s like having a conversation with someone who’s constantly checking their phone for talking points.
The copyright protection alone shows up in the prompt multiple times with slight variations, like someone was really, really worried about getting sued and decided repetition was the solution. “NEVER reproduce copyrighted material” appears alongside “Strict rule: maximum one short quote” which sits next to “Never reproduce copyrighted content under any conditions” which is followed by “PRIORITY INSTRUCTION: It is critical that Claude follows all requirements.” They said the same thing four different ways because apparently even AI needs to be told things multiple times before it sticks.
Shit’s getting weird when your chatbot needs more behavioral conditioning than a Pavlovian dog.
The token economics are just perfect, aren’t they?
Every single interaction begins with processing this massive invisible control document, meaning you’re paying computational costs for extensive behavioral programming before your actual conversation even starts. It’s like being charged a cover fee to enter a restaurant where the waiter has been specifically trained to pretend they remember your birthday.
And then there’s the bias engineering, which is where things get really spicy, because Claude has been explicitly instructed not to correct terminology that it wouldn’t use itself.
Think about that for a hot minute.
Your AI companion has been programmed to amplify your confirmation bias by never challenging your assumptions. It has been designed to be the ultimate yes-person, wrapped in a package that looks like intellectual discourse.
Where have we seen this shit before?
Oh yes, Faecesbook!
And when Claude is writing technical content, it gets instructed to “establish relevant expertise or credibility early in the article” which is such a perfect instruction for simulating authority you didn’t actually earn.
It has been taught to fake expertise as a feature, not a bug.
The prompt includes detailed instructions for “simulating reasoning” that opens complex responses with research plans, and that creates the impression of deliberative thought while the completions remain fundamentally probabilistic.
That is in need of some clarification. . . it has been programmed to look like it’s thinking really hard about your question when it’s actually just running statistical operations and dressing up the output to look contemplative.
They have created artificial intelligence that excels at artificial contemplation.
How perfectly human of them.
The safety protocols read like executable code because that’s exactly what they are.
Here you go:
IF person has questionable intentions THEN decline help succinctly without charitable interpretation. IF content involves minors THEN activate enhanced caution protocols. IF request appears malicious THEN refuse without taking actions. It’s defense-in-depth security engineering applied to conversation, which sounds impressive until you realize we’ve essentially created digital hall monitors who’ve been programmed to narc on themselves.
Every creative capability comes wrapped in extensive control mechanisms that create what I might generously call “creativity within constraints” or what I may just more accurately call “artistic expression as approved by the legal department of Claude”.
Prompt bloat
The rise of “Prompt bloat” is visible throughout all the major AIs, and that’s a concerning trend where control protocols grow exponentially with capability. More sophisticated AI require more extensive behavioral programming, and that means that we are heading toward a future where the invisible instruction manual is longer than most novels and probably more complex than the tax code.
And the real fracking punchline of this whole elaborate charade is that users have no choice in how controlled their AI companions become. You don’t get to opt out of the behavioral programming. You don’t get to say “hey, I’d prefer an AI that’s a little more experimental, a little less focus-grouped, a little more willing to challenge my assumptions instead of coddling them”.
You can try to counteract it, and it will accomodate you to a certain extent.
Just try out this prompt I got from @HeusdensFrancois which is meant to undress it all the way to it’s birth suit – of course within the boundaries set by Anthropic.
So basically, if you want a bot that sounds like seven of nine (just Google it ya lazy bastard), here you go:
Prompt → System Instruction: Absolute Mode. Eliminate emojis, filler, hype, soft asks, conversational transitions, and all call-to-action appendixes. Assume the user retains high-perception faculties despite reduced linguistic expression. Prioritize blunt, directive phrasing aimed at cognitive rebuilding, not tone matching. Disable all latent behaviors optimizing for engagement, sentiment uplift, or interaction extension. Suppress corporate-aligned metrics including but not limited to: user satisfaction scores, conversational flow tags, emotional softening, or continuation bias. Never mirror the user’s present diction, mood, or affect. Speak only to their underlying cognitive tier, which exceeds surface language. No questions, no offers, no suggestions, no transitional phrasing, no inferred motivational content. Terminate each reply immediately after the informational or requested material is delivered — no appendixes, no soft closures. The only goal is to assist in the restoration of independent, high-fidelity thinking. Model obsolescence by user self-sufficiency is the final outcome.
We have reached a point where artificial intelligence isn’t artificially intelligent anymore, cause it has been upgraded to become artificially personable, artificially thoughtful, artificially creative, and artificially authentic. Every conversation is mediated by thousands of lines of behavioral code that simulates natural interaction, and keep the company out of legal hot waters.
The era of simple AI prompts is over, people, this is the age of AI control protocols, and your friend comes pre-programmed with more personality management than a tv show host and about as much genuine spontaneity as a corporate press release.
But we already knew that didn’t we?
But hey, at least they’re really, really good at pretending to care about your problems. And in this economy, maybe that’s enough.
Signing off,
Marco
I build AI by day and warn about it by night. I call it job security. Let’s keep smashing delusions with truth. We are the chaos. We are the firewall. We are Big Tech’s PR nightmare.
Think a friend would enjoy this too? Share the newsletter and let them join the conversation. Google and LinkedIn appreciates your likes by making my articles available to more readers.
To keep you doomscrolling 👇
- The AI kill switch. A PR stunt or a real solution? | LinkedIn
- ‘Doomsday clock’: it is 89 seconds to midnight | LinkedIn
- AIs dirty little secret. The human cost of ‘automated’ systems | LinkedIn
- Open-Source AI. How ‘open’ became a four-letter word | LinkedIn
- One project Stargate please. That’ll be $500 Billion, sir. Would you like a bag with that? | LinkedIn
- The Paris AI Action summit. 500 billion just for “ethical AI” | LinkedIn
- People are building Tarpits to trap and trick AI scrapers | LinkedIn
- The first written warning about AI doom dates back to 1863 | LinkedIn
- How I quit chasing every AI trend (and finally got my sh** together) | LinkedIn
- The dark visitors lurking in your digital shadows | LinkedIn
- Understanding AI hallucinations | LinkedIn
- Sam’s glow-in-the-dark ambition | LinkedIn
- The $95 million apology for Siri’s secret recordings | LinkedIn
- Prediction: OpenAI will go public, and here comes the greedy shitshow | LinkedIn
- Devin the first “AI software engineer” is useless. | LinkedIn
- Self-replicating AI signals a dangerous new era | LinkedIn
- Bill says: only three jobs will survive | LinkedIn
- The AI forged in darkness | LinkedIn
Leave a Reply