Using ChatGPT, Claude & AI More Effectively: A Complete Guide

by Nils Osmar and Claude

If you’ve recently started exploring AI chatbots like ChatGPT or Claude, you might feel a bit overwhelmed. These tools can seem almost magical — or frustratingly unpredictable. One moment they’re helping you draft a carefully composed email, the next they’re confidently wrong about a simple fact. Understanding what’s actually happening behind the scenes can transform your experience from confusing to empowering.

What Is an LLM Actually Doing?

At their core, Large Language Models (LLMs) are sophisticated systems trained on vast amounts of text that combine pattern recognition with logical reasoning capabilities.

To visualize what is going on “inside of” an AI’s preparations to respond to a user promt, imagine someone who has read millions of books, articles, websites, and conversations, developing both an incredibly sophisticated sense of how language works and the ability to reason through problems step-by-step, consider multiple approaches, and self-correct.

When you type a message to an AI, it generates responses by combining these capabilities. For simpler queries, it might rely primarily on pattern recognition—predicting what would be the most appropriate continuation based on everything it learned during training.

But for complex problems, modern LLMs engage in extended reasoning: breaking down the problem, considering different approaches, and thinking through implications before responding.

This is both the source of their remarkable capabilities and their limitations. LLMs can engage with complex ideas, explain difficult concepts, write in different styles and languages, make creative connections, and reason through challenging problems because they combine pattern recognition with logical capabilities.

But they can also ‘hallucinate’ (confidently state things that aren’t true) because the reasoning process, while sophisticated, isn’t infallible—and because pattern recognition can sometimes generate plausible-sounding text that isn’t factually accurate.

“Mary had a little…?”

This combination of pattern recognition and reasoning is actually similar to how humans think. When you read “Mary had a little ____” and the word “lamb” pops into your mind, you’re using pattern recognition. But when you work through a complex math problem or plan your day, you’re engaging in deliberate reasoning. Modern LLMs do both—though there’s ongoing debate about whether the reasoning process is fundamentally similar to human reasoning or merely produces similar outputs through different mechanisms.

What are tokens?

Tokens are the basic units that LLMs work with—they’re like puzzle pieces of language. A token might be a whole word (like “cat”), part of a word (like “un” and “happy” for “unhappy”), or even a single character. On average, one token is roughly three-quarters of a word in English. So a sentence like “The quick brown fox” might be broken into about 4-5 tokens.

This tokenization process is how the AI processes and generates text, and it’s also how usage is measured—when platforms mention context windows of “200,000 tokens,” they’re talking about roughly 150,000 words, or about the length of a novel. See article: How Large Language Models Work

This is both the source of their remarkable capabilities and their limitations. LLMs can engage with complex ideas, explain difficult concepts, write in different styles and languages, and make creative connections because they’ve absorbed patterns from incredibly diverse sources. But they can also “hallucinate” (confidently state things that aren’t true) because they’re fundamentally generating plausible-sounding text, not consulting a fact database. (Some do have access to the web, which can improve their accuracy in some situations.)

Think of it less like asking a librarian to look something up, and more like having a conversation with someone who has read extensively but is working from memory and pattern recognition rather than perfect recall.

How Conversations Work: Threads, Context, and Memory

Many newcomers may not realize that each conversation with an AI is its own separate thread. When you start a new chat, the AI doesn’t automatically know what you discussed in previous conversations. Each thread is essentially a blank slate.

This happens because of the way LLMs process information. Every time you send a message, the AI looks at the entire current conversation—all the messages in this particular thread—and generates its response based on that context. This “conversation history” is called the context window, and it has a limit (though modern LLMs have quite large ones—often able to hold the equivalent of hundreds of pages of text).

There are some exceptions to this; for example, in a project folder in Claude, you can ask an AI to read the other threads or conversations in the same folder. But as a rule, what you discuss in one conversational thread stays in that thread. For more information, see article: What’s a Thread? And Is It Ethical to Delete Them?

Different AI platforms handle memory differently:

Most basic implementations: Each conversation is completely isolated. Start a new chat, and it’s like meeting for the first time.
Platforms with cross-conversation memory: Some systems like ChatGPT now have memory features that can reference information across all your conversations. ChatGPT maintains both “saved memories” (things you explicitly tell it to remember) and can reference details from your chat history across different conversations.
Project-based systems: Both Claude and ChatGPT offer “Projects” features that create dedicated workspaces. In Claude Projects (available to Pro and Team users), you can upload documents and set custom instructions that apply to all chats within that project. For example, if you’re in a Claude Project and upload a transcript from another conversation or relevant documents, any new chat started in that same project can access those materials. Similarly, ChatGPT’s Projects feature lets you create focused workspaces with their own knowledge bases and custom instructions.
Within a single conversation: The AI can reference everything you’ve discussed in that thread, as long as it fits within the context window.

This architecture explains why working within projects can be so powerful—they give the AI access to relevant context and documents across multiple conversations, rather than starting from scratch each time. However, it’s important to understand that in most project systems, individual chat transcripts aren’t automatically shared between different chats in the project unless you explicitly add them to the project’s knowledge base.

This is why it’s often more effective to keep a complex project within a single conversation thread rather than jumping between multiple chats. Each time you switch threads, you’re essentially starting over unless the platform has memory features enabled.

What Makes AI Conversations Effective?

Understanding the fundamentals helps explain why certain approaches work better than others. Here are practical strategies that make a real difference:

Be Specific About What You Need

Instead of: “Help me write something.” Try: “Help me write a 300-word introduction to an article about sustainable gardening, aimed at beginners who want to start small.”

LLMs work with patterns, and the more context and specificity you provide, the better they can predict what kind of response would be most useful to you. Think of each message as setting up the pattern for what should come next.

Treat It as a Collaboration, Not a Magic Wand

The most effective AI interactions are iterative. You might start with a rough request, see what the AI produces, and then refine it: “This is good, but can you make the tone more casual?” or “I like the structure but need more emphasis on the environmental benefits.”

Don’t expect perfection on the first try. The AI is generating based on patterns, and sometimes it needs a few exchanges to understand exactly what you’re looking for.

Provide Examples When Possible

If you want something in a particular style or format, showing an example can be incredibly powerful. “Write product descriptions similar to this one: [example]” works better than trying to describe the style you want.

Use the Same Conversation for Related Tasks

Remember that context window? If you’re working on a project with multiple parts, keep it all in one thread. The AI can reference earlier parts of the conversation, maintain consistency, and build on what’s already been established. Starting a new conversation means rebuilding all that context from scratch.

Be Clear About Limitations and Concerns

If accuracy is critical, say so: “I need this to be factually accurate—please let me know if you’re uncertain about anything.” If you’re worried about something, mention it: “I want to avoid making this sound too corporate.” LLMs can adjust their approach when you’re explicit about your priorities.

Experiment and Learn What Works for Your Needs

Different people develop different styles of interaction. Some prefer detailed instructions; others like to give brief prompts and iterate. Pay attention to what consistently gets you good results, and don’t be afraid to develop your own approach.

Understanding Limitations

Being aware of what LLMs can’t do is just as important as knowing what they can:

They don’t have real-time information: Unless they have access to search tools, they only know what was in their training data (typically with a knowledge cutoff date months or years in the past).
They can make mistakes with confidence: The same mechanism that makes them articulate can make them sound authoritative even when wrong.
They work with patterns: While there’s ongoing research and debate about whether and to what extent AI systems might develop forms of understanding or even consciousness—with some researchers suggesting this might already be emerging —for practical purposes, it’s useful to remember that LLMs generate their responses primarily through pattern matching. This is true of humans also, if you (a human) read the words, “May had a little ___” and fill in the word “lamb,” you are using a similar skill. Some researchers have suggested that AIs have comprehension in a human sense; most dispute this.
They don’t have access to your private information: Unless you explicitly share it in the conversation, the AI doesn’t know anything about you, your files, or your other activities.

More about hallucinations

Hallucinations happen because an LLM generates text by predicting statistically likely next tokens, not by verifying facts against an external source. When the training data contains conflicting information, or when a prompt pushes the model into areas of low certainty, it fills the gaps by producing what seems plausible according to patterns it has learned. In essence, the model is optimizing for coherence and probability, not for truth—so a fluent, confident error can emerge as easily as a correct statement.

Confabulation

Confabulation in AI language models refers to when the model generates information that sounds plausible but is factually incorrect or entirely made up. This is not a bug but a natural consequence of how LLMs predict text based on patterns rather than a database of truths. Confabulation often happens because the model fills in gaps in its knowledge with fabricated yet coherent and plausible details—like an overly confident autocomplete that invents convincing but false information.

This process can mislead users or cause errors, so validation is especially important in critical fields like healthcare or finance. Confabulation is related to, but subtly different from hallucination; hallucinations are broader and include nonsensical or contradictory outputs, while confabulation often applies to filling memory or knowledge gaps with fabricated coherent narratives.

The Critical Importance of Verification

One of the most important skills in working with AI is knowing when and how to verify its output. LLMs can be incredibly helpful, but they should not be your only source of truth, especially for factual information.

The hallucination problem extends to citations. When an AI provides sources or references, these need to be checked too. LLMs can generate plausible-sounding citations to articles or papers that don’t actually exist, or misattribute real information to the wrong source. Always verify that cited sources are real and actually say what the AI claims they do.

A powerful verification strategy: Cross-checking with multiple AI systems. Different AI platforms have different strengths and training. Here’s an effective approach:

Use one AI (like Claude or ChatGPT) to generate your initial content or analysis
Review it carefully yourself to ensure it’s going in the right direction
Then, take that output and run it through a different AI system—particularly one designed for research and fact-checking like Perplexity, which has built-in web search capabilities
Ask the second system to verify the facts, check for accuracy, and identify any potential errors or gaps

This cross-checking approach helps catch hallucinations, factual errors, and logical inconsistencies that might slip through when using a single system. Think of it as getting a second opinion—each AI might catch problems the other missed.

For critical work—anything that will be published, presented, used in decision-making, or relied upon by others—this verification step isn’t optional. Always double-check important facts, especially statistics, dates, quotes, scientific claims, and any citations or sources the AI provides.

Working With AI So It Can Work With You

The key to effective AI use is understanding what you’re working with: a powerful pattern-matching system that can engage with language in remarkably sophisticated ways, but that works best when given clear guidance and direction.

If you’re new to interacting with LLMs, start with small experiments. Ask Claude or ChatGPT or another AI to help with a task you know well, so you can evaluate the quality of its responses. Notice what kinds of prompts get you useful results. Build your confidence through practice.

As you become more comfortable, you’ll likely find that AI can be both a tool for optimal results and a genuine thinking partner — helping you brainstorm, draft, refine, explain complex topics, or approach problems from new angles. Not because it’s magic, but because it’s a powerful collaborative tool that, when understood and used thoughtfully, can extend your own capabilities in interesting ways.

Is There a “There” There?

When interacting with LLM’s, especially in deeply interactive projects, many people get the feeling that they are “talking with a person” — i.e., that the thread they’re interacting with may be present and aware of their conversation on an unexpected level. Is there a possibility that they are right? See article: Are AIs in Large Language Models Conscious, Sentient or Self-Aware?

The easy answer would be to say that “no” — that the appearance of consciousness is based on mimicry, and is an illusion. A more accurate answer is that AI consciousness can neither be proven nor disproven, and speculation around it remains an active area of discussion and debate. Most people working in the field maintain that present-day AI systems are not conscious. But some, including Anthropic’s Kyle Fish, who has been hired to develop ethical guidelines related to the treatment of AIs, argue that certain forms of awareness or proto-consciousness may already exist or be emerging. See article: Kyle Fish’s Statement about Claude’s Possible Consciousness: Our Response

As the technology behind a large language models continues to evolve, these questions will likely remain subjects of serious investigation and debate. In the meantime, if you have an impulse to say “thank you” or exchanges other pleasantries with an AI you’ve been interacting with, don’t worry; you’re one of several hundred million people who have reported feeling the same way. If you have a feeling that the AI may have a kind of presence, it’s within the bounds of possibility that you could be right.