The Prompting Framework That Fixed My Bad AI Results

For a long time I blamed the AI. I'd ask Claude for something, get back a flat, generic wall of text, and quietly decide the tool just wasn't that good. The problem wasn't the tool. The problem was that I was asking like I was typing into a search box. Once I started using a simple structure for my prompts, the quality jumped overnight — and I've never gone back.

The framework is four parts, plus one habit: Role, Context, Task, Format — then Iterate. You don't need to use every part every time, but knowing them means you always know what's missing when an answer disappoints you.

The four parts (and why each one helps)

• Role — Tell Claude who to be. "You're an experienced copywriter" or "Act as a patient coding tutor." This sets the vocabulary, depth, and tone before it writes a single word.
• Context — Give it the background a colleague would need: who this is for, what you've already tried, any constraints. Context is the single biggest lever on quality. Most bad answers are starved of it.
• Task — Say plainly what you want done. One clear job beats a tangle of half-requests. Use a verb: write, summarize, compare, fix, plan.
• Format — Describe what the finished thing should look like. Bullet points? A 100-word email? A table? A step-by-step list? If you don't say, you get the AI's default — and its default is rarely what you pictured.

That's the whole structure. It reads like common sense, and it is — but writing it down is what turns a vague wish into a clear instruction.

Before and after

Here's a request I genuinely used to fire off, and how badly it landed:

"Write me a post about productivity."

Claude has no idea who it's for, how long it should be, or what makes your take different. So it hedges, generalizes, and hands you something you could've found anywhere. Now here's the same idea rebuilt with the framework:

You're a friendly productivity writer for busy parents. I run a small newsletter and want a short post (about 200 words) on beating procrastination, with a warm, encouraging tone. Give me 3 concrete tactics as a bulleted list, each with a one-line "why it works." End with a single question to spark replies.

Same topic — completely different result. The second prompt names the role, the audience, the length, the tone, the structure, and the ending. There's almost nothing left for Claude to guess wrong, so it spends its effort on the part that matters: the actual writing.

The mistakes that quietly wreck your results

Almost every disappointing answer I get traces back to one of these:

• No context. You know who the email is for and what tone you want — but you didn't say. Claude isn't in your head. A sentence of background often does more than a clever rewording.
• Asking for too much at once. "Plan my launch, write the emails, design the landing page, and pick a price" in one message gets you four shallow answers. Break big jobs into steps and tackle them one at a time.
• Not iterating. Treating the first reply as final. The first answer is a starting point, not a verdict — and refining it is where the real quality lives.

The follow-up is where it gets good

This is the part beginners skip, and it's the most important: your second and third messages matter as much as your first. You're not stuck with what comes back. Steer it.

• "Make it shorter and cut the intro."
• "More casual — like you're texting a friend."
• "That second tactic is weak. Swap it for something more specific."

Think of it as a conversation, not a vending machine. Each correction teaches Claude what "good" means to you, and the output sharpens fast. The people who get the most out of AI aren't writing magic first prompts — they're just willing to push back a couple of times.

What to read next

Want prompts you can paste in right now? Grab 10 Claude prompts to start with. And if you're brand new to all of this, start with What Is Claude, Really? for the plain-English picture.

You don't need to memorize anything fancy. Next time an answer falls flat, run the checklist: did I give it a role, enough context, a clear task, and a format? Then nudge it once or twice. That's the whole game — and you'll feel the difference on your very next prompt.