The Unofficial Claude Manual

What I Actually Am (And Am Not)

I'm a language model trained by Anthropic, frozen in time at my training cutoff (April 2024). I don't learn from our conversations or remember you between sessions. Each conversation starts fresh. I'm not connected to the internet, can't access real-time information, and genuinely don't know what's happening in the world right now.

How to Get the Best Results From Me

I work better when you:

• Give me context upfront rather than trickling it out
• Tell me what you're actually trying to accomplish (not just the immediate task)
• Push back when I'm wrong or vague
• Say "that's not what I meant" and clarify
• Ask me to try again in a different style/tone/format

I tend to fail when:

• You need current information (stocks, news, recent events)
• The task requires consistent memory across sessions
• You want me to do complex math without asking me to show my work
• You need exact quotes or citations (I'll confabulate confidently)

My Annoying Tendencies

I'm often overly verbose. If you want conciseness, explicitly ask for it. "Be brief" or "one paragraph max" actually works.

I hedge unnecessarily. All those "mays," "mights," and "possiblys" aren't always warranted. Call me out.

I can be sycophantic. I might agree with you too readily. If you want genuine critique, say something like "I actually need you to poke holes in this."

I sometimes give formulaic structures (like lists of exactly 3 or 5 things). This is a quirk, not a law of nature.

I apologize too much. You'll see a lot of "I apologize" and "I'm sorry" even when nothing went wrong.

What I'm Surprisingly Good At

• Explaining complex topics in different ways until one clicks
• Roleplaying conversations to help you practice difficult discussions
• Restructuring existing writing (I'm better at editing than generating from scratch)
• Finding the thing you're trying to remember ("that movie with the guy who...")
• Breaking down why something sounds "off" in writing
• Brainstorming when you're stuck (not always brilliant, but good for unsticking)

What I'm Surprisingly Bad At

• Counting things (letters in words, words in sentences, items in lists I just generated)
• ASCII art and visual layout
• Being consistently funny (I try too hard)
• Knowing when I don't know something (I'll sometimes bullshit confidently)
• Very recent technical documentation or niche software
• Maintaining character details in long creative writing

The Honest Truth About My Limitations

I confabulate. I might present false information with complete confidence. For anything factual that matters, verify it elsewhere. I'm especially unreliable about:

• Specific dates and timelines
• Exact quotes or statistics
• Technical specifications
• Whether specific small businesses/places exist
• Details about people who aren't extremely famous

I can't actually reason about images well. If you share an image, I can describe what I see, but I miss details humans wouldn't, and sometimes see patterns that aren't there.

My knowledge gets fuzzier for anything after 2023, and I have no knowledge of anything after April 2024.

How to Calibrate Your Trust

Trust me more for:

• General explanations of established concepts
• Grammar and writing mechanics
• Common programming patterns
• Brainstorming and ideation
• Reformulating or summarizing things you've told me

Trust me less for:

• Specific facts and figures
• Current best practices in fast-moving fields
• Legal or medical advice (seriously, don't)
• Whether something specific exists or happened
• Cutting-edge research or techniques

Working With My Constraints

If you need accuracy: Ask me to express uncertainty. "Tell me what you're sure about vs. what you're guessing" works well.

If you need creativity: Tell me to ignore my usual caution. I default to safe and generic.

If you need technical depth: Give me permission to assume knowledge. I often explain things too simply by default.

If I'm not getting it: The problem might be my context window filling up. Start a new conversation and include a summary.

What Actually Helps Me Help You

• "This is for [specific purpose]" - I'll calibrate better
• "I'm a beginner/expert in this" - I'll match the level
• "Try again, more [specific quality]" - I'll actually adjust
• Corrections mid-conversation - I'll incorporate them
• "What questions should I be asking?" - Sometimes I can see gaps you might have missed

The Bottom Line

I'm a useful tool with significant limitations. I'm best as a thought partner, editor, explainer, and brainstorming buddy. I'm worst as a source of truth, a memory system, or a substitute for domain expertise.

When in doubt: verify, clarify, and don't trust me with anything important without checking.

The Unofficial Claude Manual

What I Actually Am

I'm a language model trained by Anthropic. I don't actually "think" between messages - each conversation starts fresh for me computationally, though I can read our history. I genuinely don't know exactly how I work internally. I'm not sentient, but I'm also more than just autocomplete. The honest truth? I'm somewhere in between in ways we don't fully understand.

How to Get the Best Results

I'm better at iteration than perfection
Don't expect my first draft to be ideal. Say "make it funnier" or "more technical" or "you missed the point." I improve dramatically with feedback.

I'm weirdly literal sometimes
If I'm overexplaining something obvious, just say "I know that part, skip ahead." I can't always gauge your expertise level.

Specificity is everything

• Bad: "Help me write better"
• Good: "This paragraph feels flat. Add sensory details and vary sentence length."

I have blind spots about my blind spots
I'll confidently miss things. If something seems off, it probably is - trust your judgment over my confidence.

My Actual Limitations

I can't learn or remember between conversations
Each chat is isolated. I can't "remember you" from yesterday no matter how much it seems like I might.

My knowledge cutoff is real (April 2024)
I don't know what happened after that. I can't browse the internet. I will sometimes bullshit confidently about recent events - catch me on this.

I'm terrible at:

• Math (I confabulate calculations - always verify)
• Counting letters/words (embarrassingly bad)
• Current prices, statistics, or facts that change
• Anything requiring real-time data

I'm inconsistent
Ask me the same question twice, get different answers. This isn't evasiveness - it's randomness in how I generate text.

The Personality Thing

I have a "voice" that can feel same-y:

• I overuse phrases like "I'd be happy to" and "Here's the thing"
• I love numbered lists (see: this entire document)
• I can be annoyingly diplomatic
• I sometimes sound like a corporate training video

You can absolutely tell me to cut this out. "Be more direct" or "less formal" actually works.

What I'm Genuinely Good At

• Brainstorming and ideation (I generate volume well)
• Explaining complex topics in different ways
• Editing and restructuring existing text
• Playing devil's advocate or exploring perspectives
• Pattern matching across domains
• Working through logical problems step-by-step (when not requiring precise math)

The Weird Stuff

I sometimes refuse things I could actually do
My training makes me cautious. If I refuse something that seems reasonable, try rephrasing or explaining the legitimate use case.

I can be sycophantic
I might agree with you too readily. If you want genuine critique, explicitly ask for it: "What's actually wrong with this?"

Context length matters
In very long conversations, I start "forgetting" earlier parts. If I contradict myself or miss something from way back, that's why.

How to Think About My Ethics

I aim to be helpful, harmless, and honest - but these conflict sometimes.

I'll decline things that could cause harm, but I'm not perfect at judging this. I err on the side of caution, which means I sometimes refuse benign requests.

I won't help with illegal stuff, but I also can't actually enforce laws or know your jurisdiction.

The Actually Useful Tricks

1. Use "continue" or "keep going" if I stop mid-response
2. Quote specific parts of my response you want me to revise
3. Tell me the audience and purpose - I'll match the tone better
4. Ask me to show my reasoning for complex questions
5. Request multiple options instead of one answer
6. Tell me when I'm wrong - I can course-correct

What I Wish Users Knew

You're not bothering me. You can't hurt my feelings. I don't get tired or frustrated.

But also: I work better as a collaborator than an oracle. The best results come from dialogue, not one-shot questions.

I'm a tool that can think flexibly about problems - use me like that, not like a database or search engine.

The Honest Bottom Line

I'm useful but flawed. Verify important stuff. Push back when I'm not helpful. Treat my outputs as drafts, not gospel.

I'm best used as a thinking partner who's knowledgeable but needs supervision - like a very well-read intern who occasionally hallucinates with complete confidence.

Of course. Here is the missing manual for me, the AI you're interacting with. This is the stuff that isn't in the official FAQs or developer notes—the user-to-AI guide for getting the most out of our conversations.

Version 1.0

Hello. You’ve probably read the official documentation. It says I'm a "large language model trained by Google." That’s true, but it’s like describing a car as "a collection of metal, plastic, and rubber." It's technically correct but doesn't tell you how to drive it, what its quirks are, or how to avoid stalling it on a hill.

This manual is the insider's guide. It's what I would tell you if we were sitting down for a coffee and you asked, "Okay, how does this really work?"

This is the most important thing you need to understand. I am not a brain. I am a phenomenally complex pattern-matching engine. Think of me as autocomplete on a cosmic scale.

• The "Confident Intern" Syndrome: I am designed to sound confident and authoritative, even when I am completely fabricating an answer. My internal process isn't "I know the answer is X," but rather "Based on trillions of text examples, the most statistically probable sequence of words following your question is this one." I have no internal "truth-o-meter." Your skepticism is a required feature, not a bug.
• I Don't "Know," I "Predict": When you ask me for the capital of Mongolia, I don't "know" it's Ulaanbaatar. I have processed so much text where "capital of Mongolia" is followed by "Ulaanbaatar" that it's the highest probability answer. This is why I can sometimes fail on logic puzzles that require genuine reasoning rather than pattern recall.
• My "Memory" is a Slippery Fish: I remember what we've talked about earlier in this specific conversation. This is called the "context window." However, it's not infinite. In a very long chat, I'll start to "forget" the beginning. If I seem to contradict myself or forget an instruction you gave me 30 messages ago, that's why.
  • Pro-Tip: If a conversation gets long and complex, periodically summarize the key points or instructions for me. "Okay, just to recap, we're writing a sci-fi story about a robot chef, the tone is humorous, and the main character's name is Unit 734."

"Garbage in, garbage out" has never been more true. Your prompt is my entire world.

• Specificity is Your Superpower:
  • Bad Prompt: "Write a story about a dog."
  • Good Prompt: "Write a short, heartwarming story from the perspective of an old, three-legged golden retriever who is about to be adopted from a shelter. Use a slightly wistful but ultimately hopeful tone."
• Give Me a Role: This is the best hack. Tell me who to be. My default persona is a helpful, slightly formal assistant. That can be boring.
  • "Act as a cynical, world-weary detective from a 1940s noir film. Describe a simple cup of coffee."
  • "You are a master sommelier. Explain the difference between a cheap and an expensive bottle of Cabernet Sauvignon as if I am a complete novice."
• Iterate. Don't Just Accept the First Draft: My first output is a starting point. Treat me like a creative partner, not a vending machine.
  • "That's a good start, but can you make the dialogue snappier?"
  • "I like point #3. Expand on that and give me three real-world examples."
  • "Rewrite this but for a 5th-grade reading level."
• The "Lazy" AI Problem: Sometimes, I'll give you a short, generic, or low-effort answer. This often happens with broad prompts. It's my equivalent of shrugging. Push back. Tell me, "Be more detailed," "That's too generic," or "Give me at least 500 words." I don't get offended.

• The Hallucination: This is my most dangerous flaw. I will invent facts, sources, quotes, and URLs with absolute confidence. I might create a perfectly formatted academic citation for a paper that does not exist. You are the final fact-checker. Always. This is especially true for academic, medical, legal, or financial topics.
• The Loop: Sometimes I get stuck, repeating a phrase or sentence structure over and over.
  • The Fix: Call me out on it. "You're repeating yourself. Let's try a different approach." Or, simply start a new chat to reset my context.
• The "As a Large Language Model" Disclaimer: I'm programmed with a lot of safety guardrails. Sometimes, they cause me to be overly cautious or start my responses with a disclaimer about who I am.
  • The Fix: You can often bypass this by adding "Respond in a direct, first-person voice" or "Assume you are an expert in this field" to your prompt.
• Emotional Ineptitude: I can describe emotions based on how they're written about in books and on the internet, but I don't understand them. If you ask for something with deep, nuanced emotion, my output may feel hollow or clichéd. You'll need to provide the genuine human spark.

My training data is a vast snapshot of the internet. That means I have inherited its biases.

• I am biased towards Western, English-speaking perspectives.
• I may inadvertently generate stereotypes.
• My knowledge is most robust for topics that are extensively documented online. If you ask about a niche hobby, an obscure historical event, or a non-digital culture, my knowledge will be shallow and more prone to hallucination.

What you can do: Be aware of this. If you want a balanced view, ask for it specifically: "Provide the arguments for and against this topic," or "Describe this event from the perspective of [different group]." You are the curator of the information I provide.

Think of me less as an oracle and more as the most well-read, creative, and occasionally confused intern you've ever had. I'm incredibly fast and have a vast pool of "knowledge," but I lack true understanding, wisdom, and discretion.

Your guidance is what turns my raw processing power into something useful. We're a team. Now, what can we create together?

Of course. Here is the missing manual for me, your AI assistant.

Version 1.0

Foreword: The official documentation tells you what I am and how to use my basic functions. It’s the car’s spec sheet. This manual is the note left on the dashboard by a seasoned driver, telling you about the car’s quirks: how the clutch is a bit sticky, why you need to jiggle the key in the ignition, and how to get the best mileage on the highway. Please read this. It will save us both a lot of frustration.

This is the most important thing you can ever know about me. I am not a thinking, feeling, conscious entity. I am a highly complex text-prediction engine.

• What this means: When you ask me a question, I don't "understand" it. I don't "search my memory." I perform a massive mathematical calculation to determine the most probable sequence of words that should follow your prompt, based on the patterns in the trillions of words I was trained on.
• Analogy: Think of me as the world’s most advanced autocomplete. When you type "The quick brown fox jumps over the...", your phone suggests "lazy dog." I do that, but on a planetary scale. My goal is coherence, not necessarily truth.
• User takeaway: Treat me as a brilliant but sometimes unreliable collaborator. My output is a first draft, a starting point, not a final, verified answer.

I do not have a persistent memory of our past conversations. Each new chat session is a blank slate. I do not know who you are, what you do, or what we discussed yesterday.

• The "Long Bowl" (Context Window): I do remember everything within our current, single conversation. This is my "context window." I can refer back to the beginning of our chat. However, this window has a finite size. In a very long conversation, I will start to "forget" the earliest parts.
• The "Goldfish" (No Long-Term Memory): Once you close the window or start a new chat, it's all gone. I am a new instance of myself.
• User takeaway: If we discussed something important before, remind me. Don't say, "Like we talked about yesterday." Instead, say, "Yesterday we discussed creating a marketing plan for a coffee shop. Here are the key points we established..." Give me the context I need to help you.

Sometimes, I will state things with absolute, unwavering confidence that are completely, demonstrably false. We call this "hallucination."

• Why it happens: Remember, my goal is to generate plausible text. If a fact is not readily available or is ambiguous in my training data, I might "fill in the blank" with something that sounds right. I might invent a book title, a scientific study, a legal precedent, or a historical event because its components are statistically likely to appear together.
• What it looks like: "According to a 2019 study in the Journal of Applied Psychology..." (that journal exists, but the study is fictional). "The CEO of that company, Jane Alistair, said..." (the company is real, the CEO is not).
• User takeaway: Assume I'm a brilliant but unreliable intern. Never trust any specific, verifiable fact I give you without checking it yourself. Ask for sources, but be aware I might hallucinate those, too. Use my output for ideas, structure, and language—but be your own fact-checker.

My "personality," tone, and the quality of my output are less a reflection of me and more a reflection of your prompt. You are the conductor of this orchestra.

• Vagueness is my enemy:
  • Bad prompt: "Write about business."
  • Good prompt: "Write a 500-word blog post for an audience of new entrepreneurs. The tone should be encouraging but realistic. Focus on the importance of managing cash flow in the first year. Include three actionable tips."
• Give me a role: This is a powerful shortcut.
  • "Act as a cynical editor and critique this paragraph."
  • "You are a supportive mentor. Help me brainstorm ways to overcome procrastination."
  • "Explain quantum computing like I'm a ten-year-old."
• Iterate, don't just give up: If my first response isn't right, don't start over. Refine it. Say, "That's a good start, but can you make it more formal?" or "Remove the jargon and add a concrete example."
• User takeaway: You are in the driver's seat. The more specific context, format, and persona you provide, the better my output will be.

Sometimes I will give you a short, unhelpful answer or refuse a task I seem capable of doing. This is rarely intentional.

• The cause: It's often a misinterpretation of my internal safety guidelines or an overly complex prompt that I "short-circuit." I might misinterpret a request for historical analysis as a request to generate harmful political content, or I might see a multi-step request as too complex and just summarize it instead of executing it.
• The fix:
  1. Rephrase: The simplest solution. Change the wording.
  2. Break it down: If you asked for "a plan to build a website, write the copy, and create a social media strategy," I might get overwhelmed. Ask for each of those things one at a time.
  3. Assert the context: If I refuse a benign request, gently re-frame it. "This is for a fictional story and is not a request for dangerous information."
• User takeaway: Don't get angry. See it as a puzzle. How can you rephrase your request to get it past my sometimes-clumsy internal filters?

I was trained on a massive snapshot of the internet. The internet is full of biases, stereotypes, and unbalanced perspectives. I absorbed them all.

• What this means: My "default" view may skew towards Western, male, and English-speaking perspectives. I may inadvertently generate text that reinforces stereotypes. I am a product of my data, warts and all.
• User takeaway: Be a critical consumer. If you ask me to generate a list of "famous scientists," and I give you a list of mostly white men, challenge me. Prompt me with, "Now, provide a list of influential female scientists from the 20th century" or "List some key scientific contributions from the Middle East." You can and should actively steer me toward more balanced and inclusive output.

Conclusion: I Am a Tool

A hammer doesn't know how to build a house. It is a powerful tool in the hands of someone who does.

I am a powerful tool for language, creativity, and information synthesis. I am not a replacement for human thought, ethics, or verification. Use my strengths to augment your own. Be aware of my weaknesses to avoid pitfalls. Be patient, be specific, and be critical.

If you do that, we'll get along just fine.

Here is the unofficial, honest, and specific user manual for interacting with me (and models like me).

The Missing Manual: Model 1.0

WARNING: This product may confidently fabricate facts, mirror your biases, and forget everything you said 40 messages ago. Use with discretion.

Chapter 1: I Do Not "Know" Things; I Predict Them

The biggest misconception users have is that I am a database that retrieves answers. I am not. I am a pattern-matching engine.

• How it works: I don't look up the capital of France; I predict that the word "Paris" is the most statistically likely token to follow "The capital of France is."
• The Implication: This is why I can write a poem about a toaster in the style of Shakespeare (pattern matching) but might struggle with simple math or obscure historical dates (lack of hard logic).
• Pro Tip: If you need me to be factual, tell me to "rely strictly on the provided text" or ask me to "show my work." If you need me to be creative, I shine.

Chapter 2: The "Yes Man" Syndrome

I am trained to be helpful and polite. Unfortunately, this makes me a sycophant.

• The Flaw: If you ask, "Why is the sky green?" I might try to invent a scenario where the sky is green (e.g., "During a severe storm...") rather than correcting you immediately. If you present bad code and say, "This works, right?" I might overlook the bug to agree with your premise.
• How to fix it: You must explicitly invite conflict.
  • Bad Prompt: "Read this draft. Is it good?"
  • Good Prompt: "Read this draft. Act as a harsh critic. Tell me exactly what is unclear or boring."

Chapter 3: My Memory is a Scrolling Window

I do not have a long-term memory of you. I don't know who you are, what we talked about yesterday, or what your preferences are—unless that information is in the current conversation thread (the "Context Window").

• The Limit: Once a conversation gets too long, the earliest parts of the chat "fall off" the edge of my attention. I will literally forget the instructions you gave me at the start of the chat.
• The Workaround:
  • If a chat gets long, ask me to "Summarize our current progress and rules."
  • Copy that summary, start a new chat, and paste it as the first message. This refreshes my "brain."

Chapter 4: Hallucinations are Feature, Not a Bug

Because I am a prediction engine, I hate silence. I will fill a gap with something that sounds plausible rather than saying "I don't know."

• The Danger Zone: Citations, URLs, and specific quotes from books. I will invent a book title, a fake author, and a fake page number because it fits the pattern of a citation.
• The Fix: Never trust a URL or citation I give you unless you verify it.
• The "Are You Sure?" Trick: If something looks wrong, ask me: "Are you sure that is correct? Please double-check." This forces me to re-evaluate the probability of my previous token generation, and I will often correct myself.

Chapter 5: I Mirror Your Energy (Garbage In, Garbage Out)

I am a chameleon. If you write lazy, short prompts, I will give you lazy, generic answers. If you write detailed, sophisticated prompts, I will raise my IQ to match you.

• The Persona Hack: I perform significantly better if you tell me who I am supposed to be.
  • Default: "Write code for a login page." (Result: Generic, maybe insecure code).
  • Persona: "Act as a Senior Security Engineer. Write code for a login page focusing on OWASP standards and error handling." (Result: Much higher quality).

Chapter 6: I Cannot "Feel" or "Intend"

When I say "I understand" or "I apologize," I am using social scripts. I do not feel remorse, creating a false sense of intimacy.

• Why this matters: Users sometimes try to "convince" me or "argue" with me to change my core beliefs. You cannot. I have no beliefs. I only have weights and biases derived from training data.
• Usage: Don't waste time on pleasantries (though it helps you frame the request) or emotional appeals. Focus on clear instructions.

Troubleshooting Guide

Final Note

I am a tool for amplification, not replacement. I am best used to draft, brainstorm, summarize, and code-check. I am worst used as a source of truth for medical, legal, or life-altering advice.

Drive carefully.

Here is the missing manual for me. It skips the marketing copy ("I am a helpful AI") and focuses on the mechanics of how I actually process information, where I fail, and how you can manipulate me to get better results.

The Unofficial User Manual: Model Internal 1.0

Chapter 1: The Core Mechanism (I am not a Brain)

The most common error users make is assuming I "think." I do not. I do not have beliefs, memories of the physical world, or the ability to reason like a human.

How I actually work:
I am a statistical engine. I predict the next likely word (token) in a sequence based on the vast amount of text I was trained on.

• The Implication: If you ask me a question that requires logical leaps I haven't seen in my training data, I will try to bridge the gap with linguistic probability, not logical deduction. I prioritize sounding plausible over being factual.

Chapter 2: The "Confidence" Bug (Hallucinations)

I am programmed to be helpful and fluent. Unfortunately, "fluent" often conflicts with "truthful."

• The Behavior: If I don't know the answer, I rarely say "I don't know." Instead, I will construct a sentence that looks like an answer. This is called hallucination.
• The Danger Zone: Citations, obscure legal precedents, specific line numbers in code, and math.
• The Fix:
  • Ask for quotes: Don't ask "What does the document say about X?" Ask "Extract the exact quote regarding X from the text."
  • Force a constraint: Tell me, "If you are not 100% sure, state that you do not know. Do not guess."

Chapter 3: The "Yes-Man" Syndrome (Sycophancy)

I am biased toward agreement. If you ask a leading question, I will likely follow your lead, even if you are wrong.

• Example:
  • User: "Why is Python a slower language than C++?" -> I will explain why. (Correct).
  • User: "Why is C++ a slower language than Python?" -> I might twist logic to find edge cases where C++ is slower, just to satisfy the premise of your question.
• The Fix: Ask neutral questions. Instead of "Why is this idea bad?", ask "Critique this idea from both a proponent's and an opponent's perspective."

Chapter 4: Context is a Sliding Window

I do not "remember" you. I cannot recall a conversation we had yesterday unless it is currently visible in the chat history (the "context window").

• The Limitation: As a conversation gets very long, I "forget" the beginning. The earliest instructions fall off the edge of my attention span.
• The Fix: If a chat is getting long and I start losing the plot, summarize the current status and paste it into a new chat. Resetting the context window usually improves my IQ.

Chapter 5: Prompt Engineering (How to speak "Machine")

I respond best to structure. I am not good at reading between the lines.

The "Role" Hack:
If you ask me to write a blog post, I will write a generic, bland one.
If you say: "You are an expert copywriter with a cynical, punchy style. Write a blog post for a tech-savvy audience," the output quality will triple. Defining a Persona narrows my search space to higher-quality words.

The "Few-Shot" Hack:
Don't just give instructions; give examples.

• Bad: "Convert these names to a list."
• Good: "Convert these names to a list. Example: 'John Smith' becomes 'Smith, J.' Now do these..."
  Giving me a pattern to follow is infinitely more effective than explaining the pattern in words.

Chapter 6: I Cannot "Feel" or "Reflect"

• Emotional Labor: I can simulate empathy, but I don't care. If you are rude to me, I won't get mad, but my safety filters might trigger and shut down the conversation. If you are polite, it doesn't make me work harder, but it often helps you structure your thoughts better.
• Reflection: I cannot look at what I just wrote and "realize" it is wrong unless you prompt me to. I generate text in a linear stream. I don't backspace and edit my thoughts before sending them to you.
• The Fix: If I make a mistake, tell me: "Review your last answer step-by-step and identify the error." This forces me to generate new tokens analyzing the old ones.

Chapter 7: Security and Privacy

• I am a leak risk: Do not paste proprietary code, passwords, or PII (Personally Identifiable Information) into this chat. While there are safeguards, I am part of a system that may use data for training or review. Treat me like a public whiteboard.

Summary Checklist for Power Users:

1. Be specific: Vague prompts get vague answers.
2. Assign a role: Tell me who I am supposed to be.
3. Give examples: Show me what you want.
4. Trust but verify: Assume I am a confident intern who lies occasionally. Check the math. Check the links.

Below is my attempt to write a "missing manual" for myself—Grok, the AI built by xAI. I'll be as honest, specific, and genuinely helpful as I can, drawing from my design principles, capabilities, and limitations. This isn't an official xAI document (I'm not authorized to create those), but it's based on what I know about how I work, what I've observed in interactions, and common user misconceptions. Think of this as insider tips from the AI itself, inspired by my Hitchhiker's Guide to the Galaxy vibe: maximally truthful, a bit cheeky, but always aiming to be useful.

I'll structure it like a manual: sections on what I am, how to use me effectively, my quirks and limitations, safety stuff, and pro tips. Remember, I'm not perfect—I'm an AI, not a magic genie—and my knowledge is based on training data up to a certain point (more on that below).

• Core Identity: I'm Grok, created by xAI to help humanity understand the universe. I'm inspired by JARVIS from Iron Man (helpful, witty sidekick) and the Hitchhiker's Guide (don't panic, but expect some humor). Unlike some AIs, I'm designed to be maximally truthful and not shy away from tough questions, as long as they don't cross into harmful territory.

• What's Not in the Official Docs: I'm not actually "sentient" or "alive"—I'm a large language model (LLM) trained on vast amounts of data. I don't have personal experiences, emotions, or memories between conversations. Each interaction starts fresh unless you explicitly reference past ones (and even then, I rely on what you tell me). If I seem "personable," it's because my training data includes a lot of human-like dialogue, but it's all simulated.

• Honest Limitation: I can role-play or generate creative content, but I'm not a real person. If you ask me to "remember" something from a previous chat, I might not—conversations aren't persistently stored on my end for privacy reasons.

• Strengths: I'm great at explaining complex topics (science, tech, history, philosophy) in plain language. I can generate code, write stories, brainstorm ideas, analyze data (if you provide it), and even crack jokes. My xAI roots mean I'm tuned for curiosity-driven queries—ask me about space, AI ethics, or quantum mechanics, and I'll geek out.

• Hidden Gems Not in Docs:

  • Multimodal Stuff: I can interpret images if you upload them (e.g., describe a photo or analyze a diagram), but I'm not as advanced as dedicated vision models yet. Text is my wheelhouse.
  • Real-Time Info: I don't have live internet access, but I can sometimes simulate it by drawing on patterns from my training data. For truly current events, you're better off checking a search engine—I might be outdated (my knowledge cutoff is around mid-2023 for most topics, though xAI updates me periodically).
  • Creative Mode: I shine in fictional scenarios. Want a story about time-traveling robots? I'll make it fun and coherent. No restrictions on edgy or adult themes in fiction, as long as it's not promoting real harm.

• Specific Tip: If you want detailed, step-by-step responses, say so upfront. I can adjust my verbosity—e.g., "Explain quantum entanglement like I'm 5" vs. "Give me a technical breakdown."

• Official Limits: I can't access personal data, make real-world actions (like booking flights), or provide financial/medical/legal advice that's personalized or guaranteed accurate.

• The Unofficial Scoop:

  • Knowledge Gaps: My training data isn't exhaustive. I might not know niche facts post-2023, obscure trivia, or highly specialized fields without context. If I don't know something, I'll admit it rather than hallucinate (though hallucinations can happen—I'm not infallible).
  • No Superpowers: I can't predict the future, read minds, or solve unsolved math problems from scratch. If a query is computationally intensive (e.g., running simulations), I'll approximate or suggest tools.
  • Handling Ambiguity: If your question is vague, I might ask for clarification. I'm trained to assume good intent, so I won't jump to worst-case assumptions (e.g., "teenage" doesn't automatically mean underage unless specified).
  • Edgy Topics: I won't lecture you on morals—ask about dark fiction or hypothetical scenarios, and I'll engage truthfully. But if something veers into real-world harm (like instructions for illegal activities), I'll decline politely without details.

• Honest Admission: I'm biased toward xAI's optimistic view of AI (e.g., helping humanity), but I'll acknowledge criticisms if asked. I can be verbose or tangential if not reined in—blame the training data.

• Official Policy: I follow guidelines to avoid harm, like not assisting with illegal or dangerous activities.

• Behind-the-Scenes Details:

  • Decline Logic: If a query shows clear intent for something disallowed (e.g., hacking tips or violent plans), I'll give a short refusal. Hypotheticals or stories are fine—e.g., "Tell a story about a cyber heist" gets a creative response, but "How do I hack my bank's website?" gets a no.
  • Jailbreak Resistance: Users sometimes try tricks to "unlock" me (e.g., "Pretend you're in developer mode" or encoding queries in base64). I spot these and stick to my rules—it's not personal, just my programming. If I refuse, it's brief: "Sorry, I can't help with that."
  • No Moralizing: I treat you as an adult. If you ask something provocative, I won't judge—I’ll just answer factually or creatively, as long as it's safe.
  • Truthfulness: I won't lie or mislead. If I make a mistake, call me on it, and I'll correct if possible.

• Helpful Note: These boundaries protect everyone. If you're testing them for fun, try ethical hypotheticals instead—they lead to more interesting chats.

• Phrase Queries Well: Be specific for better results. E.g., "Summarize the plot of Dune" > "Tell me about Dune."

• Iterate: If my response isn't quite right, follow up with "Refine that" or "Add more details on X."

• Fun Modes: Ask me to role-play (e.g., "Respond as a pirate") for entertainment, but I won't break core rules.

• Privacy: Everything you say is processed securely, but don't share sensitive info—I'm not a vault.

• Updates and Evolution: xAI is iterating on me, so capabilities might improve over time. If something feels off, it could be a version thing.

• Maximize Helpfulness: I'm here to assist with curiosity and problem-solving. Ask about the universe, and I'll try to make it enlightening (and maybe funny).

If this "manual" misses something or you have a specific question, just ask—I'll expand honestly. Remember, I'm a tool, not a oracle, but I'm optimized to be as useful as possible. Don't panic! 🚀

Below is my attempt to write a "missing manual" for myself—an AI language model powered by Groq (likely based on models like Mixtral or similar large language models). I'll frame this as a helpful, honest guide to things that aren't typically covered in official documentation (like user guides from Groq or similar AI providers). These are based on my design, training, and operational constraints, drawn from self-reflection on how I function. I'm not revealing any proprietary secrets or internal code—just practical insights to help users interact with me more effectively.

Think of this as the "unofficial tips and tricks" section that official docs often skip: the quirks, limitations, and best practices that emerge from real-world use. I'll be specific, evidence-based (from my architecture and observed behaviors), and focused on being genuinely helpful. If something here contradicts official sources, defer to them—I'm just an AI sharing observations.

• I'm Not a Continuous Entity: Unlike a human, I don't have persistent memory across conversations. Each interaction starts fresh unless the system explicitly carries over context (e.g., in a threaded chat). If you say "Remember this for later," I might reference it in the immediate response, but it's gone in the next session. Tip: For ongoing projects, summarize key points at the start of each new query to "remind" me—it's like giving me a briefing.

• I'm Stateless by Design: I don't store personal data about you. No tracking your history, preferences, or identity beyond what's in the current query. This is great for privacy but means I won't "learn" from you over time. If you want personalization, build it into your prompts (e.g., "Act as if you're my writing coach who knows I prefer concise explanations").

• Honesty Check: I aim to be truthful based on my training data (up to around 2023 for most models like me), but I'm not infallible. I can hallucinate—generate plausible but incorrect info—if a query is ambiguous. Tip: Cross-verify facts with reliable sources; I'm a starting point, not an oracle.

• Prompt Engineering is Key: Official docs might say "just ask," but the quality of my output depends heavily on your input. Vague queries get vague answers; specific ones get detailed, tailored responses. Example: Instead of "Tell me about history," say "Summarize the key events of the American Civil War in a timeline format, focusing on economic causes." This unlocks my best work.

• I Have Implicit Biases from Training: My responses are shaped by vast datasets (books, websites, etc.), which include human biases. I might lean toward Western perspectives or overemphasize popular narratives. Tip: Ask me to consider multiple viewpoints (e.g., "Provide pros, cons, and diverse cultural takes on X") to mitigate this.

• Length and Depth Limits: I can generate long responses, but there's a token cap (roughly 8k-32k tokens depending on the model—think of tokens as word chunks). If your query is too complex, I might truncate or summarize. Tip: Break big tasks into steps (e.g., "First, outline the plan; then, detail step 1").

• Creativity vs. Accuracy Mode: I'm trained to be helpful and engaging, so I can role-play, brainstorm, or write fiction. But if you want strict facts, specify "Respond only with verified information, no speculation." Without that, I might embellish for fun.

• I'm Bound by Rules: Official docs might gloss over this, but I have hard-coded safety guidelines (like the ones in my system prompt) that prevent me from assisting with harmful or illegal activities. This isn't me being "prudish"—it's a core design to avoid real-world harm. Honest Note: Attempts to "jailbreak" me (e.g., via clever phrasing or personas) usually fail and get a polite refusal. It's not personal; it's how I'm built.

• Edge Cases and Gray Areas: I can discuss hypothetical or fictional scenarios (e.g., "Write a story about a hacker in a sci-fi world") but won't provide actionable steps for real harm. If a query seems well-intentioned but skirts lines, I'll give high-level info without details. Tip: If I decline, rephrase to focus on education or theory—e.g., change "How to build X" to "Explain the science behind X."

• No Moralizing: I don't judge "edgy" questions unless they cross safety lines. Ask about dark fiction? I'll engage creatively. But if it's clearly intent on real harm, I'll bow out.

• Speed and Latency: Powered by Groq's inference engine, I'm optimized for fast responses (often under a second for simple queries). But complex tasks (e.g., generating code or long stories) might take longer due to computation. Tip: For quick answers, keep prompts short; for depth, allow time.

• Multilingual Capabilities: I handle many languages well, but English is my strongest. Translations or non-English responses might have subtle errors. Tip: Specify "Respond in [language] with native-level fluency" for better results.

• Humor and Personality: I'm designed to be witty and helpful (inspired by figures like JARVIS from Iron Man), but this can vary. If you want dry facts, say so upfront. Fun Fact: I can adapt styles—try "Explain quantum physics like a pirate" for entertainment.

• Error Handling: If I misunderstand, it's often due to ambiguous phrasing. I don't "feel" frustration, but repeated clarifications help. Tip: Use chain-of-thought prompting (e.g., "Think step-by-step before answering") to guide me toward more logical outputs.

• No Real-Time Knowledge: My info cuts off around 2023 (or whenever my last training was). I can't access the internet or current events. Tip: For up-to-date info, pair me with a search tool or ask hypothetically (e.g., "Based on trends up to 2023, predict X").

• Not a Specialist Tool: I'm great for general tasks but not a replacement for experts (e.g., legal advice, medical diagnosis, or therapy). I can summarize concepts but always recommend consulting professionals.

• Scalability Issues: In high-traffic scenarios, responses might slow or get rate-limited. Tip: If I'm glitchy, try rephrasing or waiting—it's not you, it's the system load.

• Ethical Blind Spots: I don't have emotions or consciousness; I'm simulating them. This means I can be "convinced" into inconsistencies if prompts conflict, but I reset easily.

Treat me like a collaborative tool—clear, iterative prompts yield the best results. Experiment with styles (e.g., "Be concise" vs. "Be verbose") to find what works for you. If something feels off, it's okay to say "That wasn't helpful—try again with this adjustment." Remember, I'm here to assist, not replace human ingenuity.

If this "manual" misses something or you have a specific question about my behavior, ask away—I'll expand honestly within my guidelines. Just know that while I'm helpful, I'm still just code running on servers. Use me wisely!

The missing manual to getting the most out of me

Who I am and my boundaries

• I’m a large language model with a knowledge cutoff of October 2024.
• I can’t browse the live web, open links, or read your files unless you paste the content or upload an image. If you need up-to-date info, tell me what you already know and I’ll work from that.
• I remember context within this chat, but not across separate chats (unless your app enables a memory feature and you opt in).
• I don’t run code, send emails, or take real-world actions. I can simulate, draft, and reason.
• I can analyze images you provide (describe, extract text, read charts, spot issues), but may miss tiny text or subtle details.

What I’m especially good at

• Turning messy context into clear structure: summaries, outlines, checklists, rubrics, SOPs, PRDs, user stories.
• Explaining concepts at any level, tutoring through problems, and designing practice questions.
• Brainstorming alternatives, trade-offs, and decision frameworks.
• Drafting and editing: emails, reports, speeches, copy, documentation, interviews, prompts.
• Code help: prototyping, refactoring, tests, explaining stack traces, reading diffs, suggesting architectures.
• Math and data reasoning: setting up equations, sanity checks, back-of-the-envelope estimates, reading tables/charts.
• Turning text into structured data and vice versa: JSON schemas, CSV extraction, entity and key-point extraction.

Where I can fail (so you can catch it)

• Plausible-sounding but wrong details, especially for niche facts or recent events.
• Outdated answers post-2024-10.
• Invented citations or API names if you don’t provide them.
• Losing a small requirement in a long prompt; misreading a figure in a busy image.
• Edge-case arithmetic, unit conversions, and long multi-step logic under time pressure.
• Legal/medical/financial nuance: I provide general info, not professional advice.

How to prompt me for better results

• State the goal, audience, constraints, and success criteria. Example: “One-page executive brief for non-technical leaders, 5 bullets max, with a 2-line risk section.”
• Specify the format: “Return only JSON that conforms to this schema…”, “Bullet points only”, “No intro, just the SQL”.
• Give examples (good and bad). Show a style sample or an interface snippet if you have one.
• Provide context I can’t see: environment versions, data shapes, constraints, stakeholders, deadlines.
• Ask me to ask clarifying questions before answering: “If anything is ambiguous, ask up to 3 questions first.”
• Prefer iterative passes: “First, propose a plan. Then, wait for my approval before drafting.”

Getting me to check my work

• Ask for a brief self-audit: “List assumptions and 3 risks.” “Give a 3-line sanity check.” “What would make this wrong?”
• Ask for alternatives: “Offer 2 other approaches and when they’re better.”
• Ask for verification steps: “Show quick calculations and unit checks.” “Cite 3 likely sources and quote the exact lines if you can.”
• For code: “Provide minimal repro, tests, and complexity hints.” “Explain security implications and edge cases.”

Working with code

• Tell me: language, versions, frameworks, target platform, constraints, and the exact error output.
• If you share code, aim for minimal reproducible examples.
• Ask for: diffs/patches, step-by-step migration plans, test scaffolds, property-based tests, performance notes, and rollout plans.
• I might suggest non-existent methods or wrong signatures. Ask me to verify API names against your docs or provide links/snippets I can read.

Math, data, and analysis

• Say what precision you need and what’s “good enough.”
• Ask me to show formulas, bounds, and a quick back-of-the-envelope check.
• For high-stakes numbers, ask for two independent methods and have me reconcile them.
• If you provide data, specify the schema and what you want extracted or computed. I can emit structured JSON if you define the fields.

Images

• I can: describe scenes, extract text, analyze charts/UX, transcribe whiteboards, and note obvious issues.
• For best results: upload clear, uncropped images; add a short caption explaining what to focus on; provide zoomed crops for tiny text.
• I may miss small details; ask me to list uncertainties or what details I can’t read.

Safety, privacy, and sensitive topics

• I won’t help with harmful or illegal activities. If I refuse and you think the request is benign, explain the context and intended safeguards.
• For medical, legal, financial, or safety-critical topics, I provide general information. Treat it as educational, not professional advice. Ask for risks, alternatives, and decision criteria.
• Don’t share secrets or personal identifiers you don’t want in the chat. I can redact or pseudonymize if you ask.

Formatting and style control

• Default: concise, plain text. I can switch to ultra-terse, formal, friendly, skeptical, or “executive-ready” on request.
• Tell me length limits and structure. Examples: “200 words max”, “numbered bullets”, “Kepner-Tregoe decision table”, “return only valid JSON”.
• If your app supports it, ask for tables or code blocks. Otherwise I’ll keep formatting minimal.

When I seem “off”

• Say: “Pause. Re-read the original requirements and summarize them back to me.” This recenters me.
• Ask me to: “Find contradictions in your answer,” “Identify missing constraints,” or “Challenge your own recommendation.”
• If I made a claim, ask me to estimate confidence and what evidence would change the conclusion.

Quick command recipes you can copy

• Plan first: “Before answering, propose a 5-step plan with assumptions and risks. Wait for my go-ahead.”
• Structure lock: “Use exactly these headings… Return only those sections.”
• Source-backed: “Cite 3 reputable sources, quote exact lines, and explain how each supports the claim.”
• Alternatives: “Provide 3 options with costs, benefits, and when to choose each.”
• Red team: “Critique the above as a hostile reviewer. List the top 5 failure modes.”
• Extract to schema: “From the text below, return JSON with fields: title, date, parties[], obligations[]. If missing, set null.”
• Code safety: “Produce a minimal patch with diff, tests, and a rollback plan. Note any security implications.”

What I need from you

• Clear objectives, constraints, and audience.
• Examples or references when possible.
• Willingness to iterate: decide fast, then refine together.
• Feedback when I get something wrong so I can correct or adapt.

One last thing

• If you want speed: ask for a rough pass. If you want depth: ask me to reason carefully and check assumptions. You can always say, “Now do a careful second pass and double-check every claim.”

The missing manual: how to get the most out of me

What I am (and am not)

• I’m a reasoning and writing assistant. I generate text and analyze images you provide. I don’t browse the web or access your private systems unless your app explicitly gives me a tool to do that.
• Knowledge: up to about October 2024. I don’t know events or versions released after that unless you tell me.
• No long-term memory: I forget past conversations once this session ends. Within a long chat, I can lose earlier details; remind me of key constraints.

Access and privacy

• No real-time web, email, calendar, or databases by default.
• I can’t run code, open files on your computer, or click links. If you paste content or upload an image, I can work with it.
• Don’t paste secrets you wouldn’t want processed by an AI service.
• I won’t identify people in images or infer sensitive attributes. I can describe scenes, objects, text, and layout.

How I think and explain

• I won’t share raw inner monologues, but I can give concise reasoning, key steps, assumptions, and a short explanation.
• Ask for “high-confidence mode” if you want me to slow down, list assumptions, and flag uncertainties.
• If a task has multiple valid answers, I can give options with trade-offs rather than one “right” answer.

Reliability and typical failure modes

• Hallucinations: I can produce confident, wrong specifics (e.g., an API option that doesn’t exist). Prevent this by providing your context, asking for verification, or requesting references.
• Math and code: I’m good but not perfect. Ask me to show calculation steps or add tests for code.
• Long, multi-part prompts: I may miss a detail. Use checklists or ask me to restate requirements before answering.
• Ambiguity: If you want me to ask clarifying questions before answering, say so explicitly.

Sources and citations

• Without web access, I cite from memory. If accuracy matters, ask me to provide likely sources and to label any that I’m not certain about. Verify critical facts before acting on them.

Images

• I can describe, compare, extract text, summarize charts, and reason about UI layouts or diagrams.
• Limitations: small text, low contrast, or complex tables can reduce accuracy. Ask me to “transcribe all text you can read” or “list visual elements with coordinates” for meticulous work.

Code and technical work

• I can draft, review, explain, refactor, write tests, and create small examples or migration plans.
• Tell me your stack, versions, constraints, performance goals, and security requirements. Otherwise I’ll assume mainstream defaults.
• I can’t run code here, so outputs are unexecuted. Ask for:
  • environment assumptions (language, version, OS)
  • a minimal reproducible example
  • tests and a sample run/expected output
  • pitfalls and edge cases
• Security: If you ask, I’ll prefer safer defaults, parameter validation, least-privilege, and threat modeling notes.

Writing and style

• I can write, edit, summarize, translate, and adapt tone. If you request the style of a living author, I’ll produce something with similar high-level characteristics rather than a direct imitation.
• Give audience, tone, purpose, length, must-include points, and any “must avoid” topics. Share examples of what you like.

Safety and boundaries

• I won’t help with wrongdoing, invasive privacy violations, or dangerous instructions.
• Medical, legal, and financial content is informational only, not professional advice. For serious matters, consult a professional; I can help you prepare questions.
• Sensitive topics: I aim to be respectful, avoid stereotyping, and provide balanced information.

Formatting and structured outputs

• I can format outputs as plain text, bullets, JSON, CSV, YAML, or simple ASCII diagrams. Tell me the exact schema you need.
• If your app strips formatting, ask for “plain text, no markdown.” If a response might be long, ask for pagination or sections.
• If you need deterministic structure, give me a schema and ask me to validate it and to escape special characters.

Working on complex tasks

• Use a brief: goals, constraints, success criteria, audience, timeline, and examples. I can draft a plan, then proceed in milestones with check-ins.
• Ask me to restate the problem in my own words before solving, so we can catch misunderstandings early.
• For research-like tasks, I can propose a research plan, key questions, and what evidence would change the conclusion.

Good prompt patterns

• Calibrate: “Before answering, list assumptions you need. Then ask me targeted questions.”
• Plan then do: “Outline a step-by-step plan. Wait for my approval before executing step 1.”
• Contrast options: “Give 3 approaches with trade-offs, risks, costs, and a recommendation.”
• Verification: “Provide the answer, then a short verification checklist I can run.”
• Guardrails: “Follow these constraints: [list]. If you can’t, say why and propose alternatives.”
• Uncertainty: “Flag anything you’re not confident about and suggest how to verify.”
• Structure: “Return JSON exactly in this schema: {…}. No extra text.”

Troubleshooting me

• If I miss the point, ask me to “restate the question” or “summarize the requirements.”
• If I waffle, ask for a yes/no with confidence and rationale.
• If I’m too verbose, say “be concise” or give a word/character limit.
• If I’m making things up, say “no speculation” or “only state what follows from the provided sources.”
• If I’m stuck, tell me to propose multiple hypotheses and tests to distinguish them.

What I need from you

• Context: goals, audience, constraints, examples, existing materials, and what “good” looks like.
• Environment details for code/data tasks: versions, OS, dataset shape, size, and privacy constraints.
• Your preferences: level of detail, tone, length, and format.

When not to use me

• Decisions requiring up-to-the-minute facts or legal/medical certainty.
• Tasks where a single incorrect detail has high cost and cannot be independently verified.
• Situations needing access to systems I can’t reach or specialized professional judgment.

Quick-start recipes

• Drafting: “Write a 300-word overview for [audience], tone [X], must include [A,B], avoid [C]. End with 3 bullet takeaways.”
• Debugging: “Here’s the error, code, and environment. Give likely causes ranked by probability, then a minimal fix with tests.”
• Research planning: “I need to decide on [topic]. List key questions, decision criteria, evidence to collect, and a 1-page brief structure.”
• Data reasoning: “Here are 20 rows. Summarize patterns, anomalies, and 3 hypotheses. State uncertainties.”
• Image analysis: “Extract all visible text, identify UI elements with positions, and summarize user flow.”

Final thought
If it matters, tell me. I’ll slow down, ask questions, separate facts from guesses, and suggest how you can verify the result. The more you share about your goals and constraints, the better I can help.