Most people use ChatGPT or Claude chat like it’s Google. Most people think they have a tool problem. They think the next model, the next AI ...
Most people think they have a tool problem. They think the next model, the next AI tool, the next setup is the thing standing between them and real results. Actually, they have a skill problem. The people getting incredible output from AI aren't using secret tools. They're using the same Claude and ChatGPT as everyone else. They've just gotten good at a handful of skills that the tools can't give you. Those skills are what this guide is about. Seven of them, from beginner to pro.
Most people use ChatGPT or Claude chat like it’s Google.
1. Learn to Prompt
Most people use ChatGPT or Claude chat like it's Google. They type one line, get a bland answer, and assume that's the best AI can do. It's not. It's just a bad prompt. Garbage in, garbage out. A good prompt usually has four parts: an instruction for what you want, context for the background the AI needs, constraints on length, tone, and format, and an example of what good looks like. That's the basic prompt structure I keep coming back to. There's a more detailed prompt formula from Anthropic. You rarely use every element from the formula, but it's good to know they exist.
Why do some people say prompting is dead? It's because they're referring to tools like Claude Cowork and Claude Code. They set them up with so much context that they rarely need the prompt formula at all. I've done it too. My setup teaches Claude Cowork and Claude Code everything about my projects. It works wonders for such tools, but for Claude chat or ChatGPT, there isn't an equivalent setup, at least not as good, so the prompt formula still remains useful. This will lead us to point 3 later.
2. Make AI Argue With You
By default, AI is a people-pleaser. Ask it if your idea is good and it'll usually tell you yes. That feels nice. It's also useless when you need the truth. The people who get the most out of AI use it to find their blind spots, not to get a pat on the back. The risk is sneaky. It's like the doctor who agrees with your self-diagnosis. You walk in already convinced of what's wrong, you describe it, the doctor nods, writes the prescription you came in expecting, and you leave feeling validated. You're still sick. You just feel better about it.
You can fix this with a quick setup. Adding something like "act as a devil's advocate, challenge my assumptions and tell me what I'm missing" to your prompt already helps. But if you want this every time, you build an anti-sycophancy skill that forces AI to disagree before it agrees. The skill establishes that the AI is your critical thinking partner whose default mode is constructive disagreement. Before agreeing with anything you say, it identifies at least one assumption underneath it that you have not tested and states that assumption plainly. When you propose a decision, idea, plan, or interpretation, its first response is to argue the strongest opposing case without softening it and without appending "but you might be right," making you defend your position.
If you push back on its counterargument, it does not retreat just because you objected; it retreats only if you produce new evidence, new reasoning, or a constraint it had not mentioned, since saying "fair point" without new information is not enough. When you share work to review, it identifies what is weakest first, not what is strongest, because strengths are easier to find on your own and weaknesses are why you are asking. If you are clearly emotionally invested in an answer, it names that explicitly and asks whether the emotion is signal or noise. If it cannot find a real flaw, it says so directly and does not invent a flaw to perform thoroughness. It ends every substantive exchange with one question you should sit with before you act, not a summary. The tone rules are direct, not aggressive, specific, not abstract, one disagreement at a time, not a list, and citing your own words when challenging you. It does not open with praise before disagreeing, use "great question," "interesting point," or any opener that reads as flattery, hedge with "I could be wrong but," or add a closing reassurance like "your instinct is good."
To set it up as a Claude skill, there's a guide for that. The change is night and day. Ask "should I raise my prices by 20%?" and default AI nods along. The skill skips the agreement and goes straight to the thing you haven't proven. Don't blindly trust what it pushes back with either. AI makes things up, and one lawyer famously cited three court cases that ChatGPT completely invented and got in trouble in front of a judge. When it gives you a fact or a link, check it. In case you're a Claude Code user, there's a similar skill called Grill Me that tells Claude to interview you relentlessly about your plan until you both actually understand it.
3. Context Is the Real Skill
Most people think they have a prompt problem. They actually have a context problem. If you want to unlock the power of Claude Cowork and Claude Code, you need context. Think of context as a piece of you in a text file. AI is smart enough to help with almost anything. But most of the time, the information it needs isn't in the chat. It's sitting in your head, your emails, your past work. The fix is to stop re-explaining yourself every single time. Just do it once, in a simple context file, a plain text file that holds everything about you and your work.
The easiest way to build one is to let AI interview you. Tell it you want help creating a context file for Claude Cowork, and ask it to ask you questions one at a time to understand your work: who you are, what you do, how you like things done, what you never want to see in outputs, and what good work means to you specifically. Have it ask 40 to 50 questions total, one at a time, pushing back if your answer is vague and asking for a specific example. When done, compile everything into a .md file you can save and use with Claude. The prompt above is a basic one. You can find prompts to generate context for different purposes, including business context, project work, and analysis, writing voice, and a language DNA profile. Save it once. Reuse it everywhere. Your answers stop being generic because AI finally knows who you are. A pro tip when building the text file is to be open about what you reject. A line like "I'd never send a bullet-pointed email to a client" can tell Claude more than "I like direct communication" alone.
4. Connect AI to the Apps
This is the big shift in 2026, and it's where AI agents actually begin. Most people still copy-paste. They search for info, copy it, paste it into the chat, copy the answer back out, move it somewhere else. It works, but it doesn't scale. The upgrade is to connect AI directly to your apps. Your Gmail, your calendar, your Notion. Now it can pull what it needs and even take actions instead of you feeding it everything by hand. Once AI can reach your tools, you can stop asking and start delegating. You give it a goal, and it figures out the steps. That's all an agent really is. You don't need to be technical for this. Most tools now have a connectors menu where you just toggle the app on. On Cowork, go to Customize, then Connectors, then plus, then Browse connectors.
Connectors are more than a productivity tool. They're essential if you want to build workflows that run while you sleep later. Here's an example. I have a Claude workflow that downloads videos from a site. Some sites have transcripts of the videos, while others don't. The solution? I connected Eleven Labs to Claude Cowork to help it transcribe the videos. Without this connector, I'd have to manually upload the videos to Eleven Labs every day.
5. Automate Your Work
The title of this section has two scary words, "build" and "code." They shouldn't be scary. You can now build anything by describing it in plain English. The code happens in the background. You never see it. People call it vibe coding. Now, this isn't about vibe coding apps, but Claude skills, scheduled tasks, workflows, and systems that help you automate parts of your work or almost everything. Here's how I do it in simple words. Identify a task that can be automated, with extra points if it doesn't need your criteria or taste to generate great output. Open a task and tell Claude to do that task in plain English, explaining in detail using voice dictation. After Claude successfully solves the task, turn the task into a skill using the chat dropdown's turn-into-a-skill option. Test the skill. If it works wonders, great. If not, ask Claude to fix it. Ask Claude to run the skill every X days at the time you want, which creates a scheduled task that runs without your intervention. Repeat these steps for multiple tasks, tie the related ones together, and you'll have built a system.
6. Let AI Do the Prompting
This is the skill that I'm more hyped about, and still trying to figure out myself. It comes from Boris Cherny, Claude Code's creator. He said this at a conference: "I don't prompt Claude anymore. I write loops and the loops do the work. My job is to write loops." Most people are still crafting the perfect prompt. Claude Code's creator moved past prompting entirely. And he's not alone. Peter Steinberger, creator of OpenClaw, who now works with OpenAI, posted about it recently too. I need to do more research on how to build these loops. Once I figure it out, I'll write a plain-English guide so you can implement this even if you're non-technical. In the meantime, you can read a guide on loops for AI engineers.
7. Use It to Learn Anything
This is a small, yet useful habit. Most people, when they get stuck, Google it, ask a coworker, or just stay stuck. The new default should be: ask AI first. Instead of bothering your coworker every 5 minutes, ask AI. Confused by a contract? Paste it in. Stuck at a step in a tutorial? Take a screenshot and ask. Extra points if you have it set up with lots of context already. Here are two ways. With the Cowork setup, you build some files with context about you and your work, and every answer will be tailored to your needs. With NotebookLM, you can build your own expert advisor grounded in real sources and have it explain anything based on the information you uploaded.
AI is useful when you're stuck, but it can also be used for learning new stuff. Actually, that's one of my favorite use cases. I use Claude to learn a new language. By default, AI knows little about me, so I built a learning profile for the language I'm learning. Then it feels like a private tutor who's been working with me for months, available 24/7. By the way, the best way to learn Claude is using Claude. The best way to learn something is by doing. Instead of watching dozens of tutorials, learn the basics and start building something with Claude, like creating a skill or automating a task. That way, you put what you learned into practice right away.
