I spent a lot of my damn time trying to make ChatGPT do things.
I’d write a prompt, get back something generic and cheerful, rewrite the prompt, get back something generic and cheerful in a different font. Eventually I switched to Claude, started giving it proper prompts, and the difference was obvious. That’s my opinion: for thinking through a hard problem, Claude is better at this than ChatGPT.
I spent a lot of my damn time trying to make ChatGPT do things.
Yana Semova, an IT project manager and AI trainer, has one habit that fixes a lot of muddled output: start a new project when you start something new. “Start on blank, and the result it delivers is going to be much more precise, clear and specific,” she said on Leaders and Their Stories. I keep one long project for its context and throw away fresh ones for jobs that need none, so it depends on the goal.
The tool was only half of it, though. The other half was learning how to feed it. Type a one-line question and you’ll still get fortune-cookie advice, from any model. Claude can’t see inside your business, your book or your head. It only knows what you tell it.
I use AI every day as a digital assistant.
It’s a lot like having a virtual assistant, only digital, and I still do the writing and the thinking. On a stubborn problem, though, it’s the best sparring partner I’ve ever had. Here’s the sequence I use, with two problems running through it: one of mine, and one every novelist knows.
Why do most Claude prompts give generic answers?
Because the prompt is generic. A model answers the question in front of it, and “how do I get more clients” has ten thousand reasonable answers, so you get the average of all of them. That average is the blandest, safest advice on earth. Is it the tool’s fault? Hardly, and it drives me up the wall when people blame it.
The other thing people get wrong is treating each prompt as a one-shot. These six work because they run in one conversation, in order. Each answer becomes context for the next question, so by step five Claude knows your situation better than most people you’d ask for advice.
Start with the real details
Before any of the six, open the conversation with something like this:
Here’s the situation. [What’s happening.] I’ve tried [what you tried], and it didn’t work because [why]. It’s costing me [time, money, sleep]. I’ve ruled out [options] because [reasons]. Keep all of this in mind for the next few questions.
Don’t polish it. Messy and honest beats neat and vague every single time.
My problem: I’ve got about 16,000 LinkedIn connections, I’ve worked through roughly half of them, and I want to reach the ones I’ve never messaged. LinkedIn has no button for that. Not in the regular search, not in Sales Navigator, nowhere. I’d been fighting LinkedIn’s search for months and getting nowhere.
The novelist’s problem: say you’re 40,000 words into a mystery and you’ve been stuck in the middle for six weeks. You open the file every morning, reread the last chapter, fiddle with a sentence and close it. You’ve tried outlining ahead and writing a different scene, and neither stuck.
How do you use Claude to find the root cause of a problem?
Make it interview you. The cause of a stubborn problem is sitting in your own head, in the part you skip over because it’s uncomfortable or so familiar you’ve stopped seeing it.
I keep running into this. Don’t give me solutions yet. Ask me questions, one at a time, until we find what’s causing it underneath the symptom I keep treating.
One question at a time matters. A list of twelve is easy to skim and half-answer. One at a time forces you to think, and by the fourth or fifth one of them stings.
For my LinkedIn problem, the questions get past “LinkedIn’s search is bad” to the underlying issue: I don’t have a list of who I’ve contacted anywhere I can sort it. For the novelist, they end up somewhere uncomfortable, like “what happens next?” getting a shrug, because the writer never decided what the killer wants.
How do you break a big problem into smaller pieces with Claude?
Big problems feel paralyzing because they look like one enormous thing. Break one open and you’ll find a stack of small things, with one of them holding up the rest.
Break this into the smallest pieces you can. Then tell me which single piece is blocking the others, and why.
The second sentence is the valuable half. I can make a list of pieces on my own, and so can you. What I want from the machine is the blunt call on which piece to hit first, because I’m too close to see it.
My pieces: who I’ve messaged, who I haven’t, who’s worth messaging, and what to say. The blocker is the first one, since without it the rest is guesswork.
For the novelist, the blocker is rarely the stuck scene. It’s the villain’s motive, decided badly or never decided at all, holding up every scene after it.
This is exactly how I think novelists should use these tools. My 40-plus writing handbooks are built on AI prompts that help you work out a protagonist, a villain’s motive or a plot that holds together, and none of them ask the machine to write a word of the book. Let it question you about your villain, and keep your own hands on the sentences. I also have it play the target reader. Will people read this book top to bottom, or use it as a reference? Pretend you’re the reader: where do you get bored? It’ll tell you, and then you look at what makes that spot boring.
How do you get Claude to tear your plan apart?
You have to ask bluntly, because these models lean agreeable. Ask whether your plan has problems and you’ll get three gentle “considerations” wrapped in praise. That’s worthless when you’re about to bet time or money on something. I’ve written about how the companies keep swinging the agreeableness dial back and forth, and you can’t rely on where it sits today.
Here’s how I’m handling this right now. Skip the balanced review. Give me the strongest case against this approach, the argument that would sink it if it’s true.
It stings when it hits something real, and it should. Better to feel it in a chat window than in your bank account.
The same trick works on the machine’s own output. A writer in an AI authors’ group asks for marketing copy or story beats, then tells the model to reread and critique what it just gave her. It tears its own draft apart, and the second version comes out stronger every time. Another member had the reason: the model can’t read and measure what it writes while it’s writing it. It predicts one word after the next, and the review only happens when you ask for one.
I take that a step further with code. I move it to a fresh session with no context and ask for an audit there, because the session that wrote it is carrying all the context that produced the mistakes.
My plan was to keep searching LinkedIn by hand, and the case against it is brutal: at the rate I was going, I’d finish sometime next year. The novelist’s plan was to push through by writing more scenes, and the case against that is worse. Scenes written without knowing the villain’s motive get thrown out later, so every word written now is borrowed against a rewrite.
Change the assumption, change the answer
Most of the time we’re stuck because we’ve assumed something without noticing, and every solution we try inherits the assumption. Swap it and a whole new set of answers shows up.
Give me three completely different ways to approach this, each built on a different assumption about what’s really going on. Name the assumption behind each one.
Naming the assumption is the trick. Once you see three premises side by side, you’ll know in your gut which one is true, and that’s worth more than any of the three plans.
That step broke my problem open. I’d been assuming the answer had to live inside LinkedIn. It doesn’t. LinkedIn lets you download your own connections and your message history, and once those two files are on your own computer, finding everybody you’ve never contacted is a five-minute sort. The problem stopped being about LinkedIn’s lousy search the moment the assumption changed, and I could’ve kicked myself for not seeing it sooner.
For the novelist, the assumptions might be “I don’t know my villain,” “this book has the wrong structure” and “I’m burned out and it isn’t the book at all.” Each one points to a completely different fix, and only one of them is true.
Ask the question you haven’t asked yet
This is the one people use too early. Ask it cold and you get the kind of question printed on motivational posters. Ask it after the first four steps and Claude has enough to spot the gap in what you’ve told it.
Based on everything we’ve covered, what’s a question I haven’t asked myself that might change the answer?
For my LinkedIn problem, the question is simple and a little embarrassing: once I have the list, who on it is worth my time at all? For the novelist it might be “do you still want to write this book?” That one hurts, and it’s worth a straight answer before another six weeks go by.
My take
People ask their AI one vague question, get a vague answer, and blame the tool. Feed it the real mess first. The step I’d never skip is the prompt that names your assumptions. My LinkedIn problem came apart at that step, once I saw the answer was never inside LinkedIn. Run the prompts in order, in one conversation, and don’t stop until you’ve got one move you’ll make in the next 24 hours.
What should you do after Claude helps you think through a problem?
Move, within a day. Insight that doesn’t turn into action evaporates by the weekend, and I’ve watched plenty of clever plans die in a notes app.
Based on all of this, give me one specific action I can take in the next 24 hours. Make it concrete enough to put on my calendar.
Pick one action. Five turns into a to-do list you’ll dodge for a week, and one is something you’ll do before lunch.
Mine is to request the data export today. The novelist’s might be to write one page, in the villain’s voice, explaining what they want and why they’re willing to kill for it. Neither one solves the whole problem. Both get you moving again.
More of my thinking on where these tools help and where they don’t is in the AI and Writing Hub, including why most writing prompts fail and how I split the work between me and the machine on a book. If you want help setting up AI so it works for you without flattening your voice, that’s what my AI services are for.
None of this works if you lie to it. Tell the machine the real mess, let it ask the questions you’ve been avoiding in an order that won’t let you dodge, and then go do the one thing it helped you find. People who keep asking for easy answers to problems they haven’t explained will keep getting fortune cookies, and they’ll keep blaming the tool for it.
Good prompts work because you keep doing the thinking. The Birth of the Augmented Human is my book on using AI as a force multiplier for your own judgment, so it amplifies your work instead of replacing it.
