How to write prompts that get better results from Claude

Most people blame the model when a Claude answer comes back flat. Usually the prompt was too thin. Claude does not know who you are, what you are trying to do, or what “good” looks like for you, and it will not guess well. This article shows how to write prompts that get better results from Claude: a three-part structure, a before-and-after example, how to iterate on the reply, and how to fix an answer that misses. For the model, effort, and thinking menu, see the previous article, Your first conversation with Claude on the web.

The three-part prompt structure

When you know what you want, the fastest structure is three moves. Anthropic’s Claude Academy lesson “Your first conversation with Claude” frames it as three questions to ask before you write:

1. Set the stage. What is your role and what are you trying to achieve? What context about your work should Claude know? Without this, Claude guesses, and it guesses generically.
2. Define the task. What action should Claude take: write, analyze, build, summarize, or compare?
3. Specify the rules. Which style or tone should it use, and what format and length do you want? Attach an example if you have one.

The framework is adapted from the 4D Framework for AI Fluency, developed through research collaboration between Professor Rick Dakan (Ringling College of Art and Design) and Professor Joseph Feller (University College Cork). Anthropic’s Help Center keeps the same instinct in four words: start simple, be specific, iterate, explore.

Before and after: how to write prompts that work

Here is a weak prompt:

Research the streaming market.

It has no role, no goal, no constraints, and no format — Claude has to guess all four. Here is the rewrite from the Academy course:

I'm the marketing lead at an indie streaming startup, and we're preparing an investor pitch deck for Series A investors. Can you research the current state of the independent film streaming market and identify key trends, competitor positioning, and growth opportunities? Use current web research with citations and structure it as a professional report of up to 5 pages, with an executive summary, market analysis, competitive landscape, and growth opportunities.

Four things changed:

• The stage: marketing lead at an indie streaming startup, preparing a Series A pitch deck.
• The task: research the market and identify trends, competitor positioning, and growth opportunities.
• The rules: use current web research with citations; a professional report of up to 5 pages with a named structure.
• The deliverable: an investor-ready report, not a blog post.

The point is not that this prompt is long. It is that every sentence carries a decision Claude cannot make for you.

Iterate instead of restarting

Anthropic’s lesson is blunt: “One of the most important shifts when working with Claude is recognizing that your first prompt rarely produces a perfect result—and that’s okay. Think of your initial prompt as the start of a conversation, not a one-shot request.”

Three ways to steer the conversation:

1. Ask a follow-up. “Can you expand on the second point?” / “That’s helpful, but can you make it more concise?”
2. Give feedback. “This is good, but the tone is too formal. Can you make it more conversational?”
3. Redirect or restart. “Actually, I was asking about X, not Y.” If it has gone badly off track, start a new chat.

Specific feedback beats vague feedback. “Cut the first two paragraphs and make the conclusion more action-oriented” tells Claude exactly what to do; “make it shorter” leaves it guessing. And two kinds of edit are worth knowing:

• Edit your own prompt. Click the pencil icon on any of your messages to edit and resubmit it. This is useful when you want to refine the request rather than add a new message. Anthropic’s Academy course describes this; no matching Help Center article was found, so treat the button location as approximate.
• Edit Claude’s output. In a doc or Markdown artifact, highlight the text you want changed, click “Edit with Claude,” and type your request. Claude makes the edit where you marked it, so you do not have to describe which section you mean.

Fix an answer that misses

When an answer comes back wrong, the cause is almost always one of five things:

What went wrong	What’s happening	Try this
Response too generic	Not enough context about your situation	Add audience, role, and constraints
Too long or too short	Claude is guessing at length	“Give me a two-paragraph summary”
Didn’t follow my format	Claude understood what, not how	Provide an example or describe the structure
Confident but wrong	Plausible but incorrect, especially with niche facts	Verify independently; ask for citations; enable web search
Tone isn’t right	Claude defaults to helpful and professional	Name the tone and give a sample

For high-stakes work, verify key facts independently — ask Claude to cite sources or indicate a confidence level, and enable web search to ground responses in current information.

Before you spend more tokens, check the cheap fixes first:

• Batch related questions. “If you have multiple related tasks or questions, group them in a single message.”
• Review the prompt before sending. A missing constraint is cheaper to add now than to iterate around later.
• Check your usage. Every retry costs usage. Settings > Usage shows where you stand, and the last article in this series explains how the limits work.

AI Fluency: the habit that makes the rest stick

Prompting is one of four skills in the 4D Framework for AI Fluency that Anthropic’s course draws on:

• Delegation — who does what.
• Description — communicating with AI.
• Discernment — critically evaluating outputs.
• Diligence — using AI responsibly.

This article is mostly Description. Discernment is the part beginners skip: read the answer before you trust it, and check the facts that matter.

The course also recommends a small test suite: gather 5–10 examples of a task you do regularly, write test prompts, compare the outputs, and refine your approach. The habit is simple — keep the prompts that worked. A prompt that produced a good investor report is a template for the next one.

Frequently asked questions

Does a better model fix a bad prompt?

No. Opus gives a more thorough answer to a vague request, but it still has to guess your role, goal, and format. The three-part structure matters more for most everyday work.

Can I edit a message after I have sent it?

Anthropic’s Claude Academy says you can click the pencil icon on any of your messages to edit and resubmit the prompt. We found no matching Help Center article, so treat the exact button location as approximate and check what you see on your own screen.

Does a prompt have to be long?

No. The length is not the point. Every sentence in a good prompt carries a decision Claude cannot make for you: who you are, what you want, and what the result should look like. A short prompt that covers all three beats a long one that covers none.

When should I start a new chat instead of iterating?

When the conversation has gone badly off track. Redirecting works for a small miss (“Actually, I was asking about X, not Y”). If the thread keeps drifting, a fresh chat with a better first prompt is usually cheaper than rescuing the old one.

Conclusion

Better results come from better instructions, not from a better model. Set the stage, define the task, specify the rules — then iterate with specific feedback instead of restarting. When an answer misses, add context: audience, role, constraints, and the format you want. The prompts that work are worth keeping, because the next one starts from there. The next article in this series, Uploading files and images to Claude, shows how to hand Claude your own documents, spreadsheets, and photos. You can browse every part of the series on the Claude on the web: a beginner’s guide page.