How I Actually Use Claude
Not prompts as magic words — Claude as a thinking partner you direct with intent. Here is the loop I keep coming back to.
Most people ask Claude for an answer. I learned to direct it like a collaborator — and the work that comes back is a different thing entirely.
When I started, I treated Claude like an oracle. I'd type a vague wish and hope brilliance came back. Sometimes it did, mostly it didn't, and I quietly blamed the model. The shift that changed everything was small: I stopped wishing and started directing. I tell it what I'm trying to do, who it's for, what to avoid, and what I already know — the same briefing I'd give a sharp colleague joining a project halfway through. Watch the difference on the simplest possible request:
The model was always capable of the fuller answer. It just couldn't read my mind about which one I wanted. Here's the part most people get backwards: the trap is thinking a better answer comes from a cleverer model. Most days it comes from a clearer instruction. The model in front of you is already capable of the work you want — the bottleneck is almost always how much of the situation you've actually told it.
The loop I keep coming back to
Directing isn't one good prompt; it's a loop. I direct, I read what comes back, I correct the one thing that was off, and I go again. Most of my best results are three turns of that, not one lucky shot. The whole course is really about getting good at each part of this circle — and you'll see it return, larger, when Claude starts using tools and working in your terminal.
When it disappoints you
When the output isn't what I wanted, my first move is no longer to retry the same prompt and hope for a better roll. It's to ask: what did I fail to say? Did I name the audience? Show an example of good? Mention the constraint that actually matters? Nine times out of ten the fix is a sentence I left out, not a flaw in the model. Treating a weak answer as missing information rather than bad luck is the whole craft, and everything else here is a way of getting better at it.
Where it breaks
Re-rolling the same vague prompt, expecting a different result. If you didn't change the instruction, you're not directing — you're gambling. The model can't supply the context you're holding in your head; only you can put it on the page.
Try it yourself
Take a vague request you'd normally type and add three things before you send it: who the result is for, one example of what “good” looks like, and the one thing it must avoid. Send both versions and compare. The gap between them is the whole lesson — and the reason the rest of this course is worth your time.
Grounded in Anthropic's prompt-engineering guidance — precision in, quality out.