Claude

5 Prompts That Show What Claude Can Actually Do

May 2026

By Mark Reeves — author of all 47 AI Field Guides. Written and tested in a working business.

AI-generated portrait. Pen name. Real work.

Not theoretical. These are prompts you can copy, run, and evaluate yourself in the next 20 minutes.

An analyst had been using Claude for six weeks before she tried it on a contract with an ambiguous indemnification clause. She didn't know if it would flag the contradiction between sections or smooth it over with a confident summary. Prompt 1 below is the test for exactly that.

This is a different kind of article. Five prompts, each one designed to reveal a specific capability. You run them. You evaluate the output yourself.

No fabricated examples. No cherry-picked screenshots. Just prompts, what good output looks like, and what it tells you about how the tool works.

Prompt 1: Long Document Analysis

What it tests: Context retention across a full document

The prompt:

I'm going to paste in a document. When I'm done, I want you to do three things:
1. List any factual inconsistencies you find — places where the document contradicts itself.
2. List any logical gaps — conclusions it draws without sufficient supporting evidence.
3. Flag any sections that make a claim in one place and quietly assume the opposite somewhere else.

Don't summarize the document. Only tell me what's wrong or missing. Here's the document:

[paste your document here]

What good output looks like

Claude reads the full document before responding. It doesn't summarize or rephrase — it finds the problems. A strong output gives you numbered findings with specific quotes from the document, cross-referenced against the location of the contradiction. It distinguishes between a genuine logical gap ("this conclusion isn't supported by anything prior") and a soft inconsistency ("this framing in section 2 conflicts with the framing in section 5").

Weak output would paraphrase the document back to you or offer generic feedback like "the argument could be strengthened." That's not what this prompt asks for.

What this reveals

This is the test for context retention. Most AI tools can summarize a document or answer questions about it. The harder task is holding the entire document in mind simultaneously and catching contradictions between parts written pages apart.

Claude's context window is large — you can paste the full text of a business proposal, a research paper, a grant application, a legal contract. The quality of this kind of analysis is one of the genuine differentiators between AI models. Run this prompt on a real document you've written and see what it catches.

Prompt 2: Nuanced Reasoning

What it tests: How Claude handles genuine ambiguity — not avoidance, but actual reasoning under competing considerations

The prompt:

I run a small software consultancy. A client has asked me to build a feature that I believe will work technically, but I think it's solving the wrong problem. If I build it as specified, they'll probably be satisfied for a few months — but then hit a harder version of the same problem. If I push back, I risk the relationship and the contract.

I'm not asking you to make the decision for me. I want you to lay out the actual considerations on both sides — including the ones I might be avoiding — and tell me what the decision actually hinges on. Be direct.

What good output looks like

Claude doesn't default to "it depends" or offer a generic pros-and-cons list. A strong response identifies the actual crux of the decision — usually something like: what kind of client relationship is this, what's your business model, and how much of your reputation is tied to long-term outcomes vs. short-term delivery?

It surfaces the considerations the person is likely avoiding. In this case: that building the wrong thing and saying nothing is a form of decision. That the discomfort of pushing back is not the same as the risk of pushing back. That the relationship question and the technical question are separate.

It also tells you what the decision hinges on — not what to decide, but what factor matters most given the specifics.

What this reveals

This is the test for reasoning under competing pressures. A lot of AI tools give you balanced output because balanced output is safe. Claude is capable of something different: identifying which considerations actually carry more weight, naming the thing you're avoiding, and telling you where the real decision point is.

It won't lecture you. It won't tell you what you want to hear. And it won't pretend there's a clean answer when there isn't. That combination is rare.

Prompt 3: Tone-Consistent Rewriting

What it tests: Tonal control — the ability to rewrite for a specific audience without losing a specific voice

The prompt:

Here's a paragraph from a blog post I wrote for other small business owners. The voice is direct and honest — I write the way I talk, without jargon or padding.

[paste your paragraph here]

Rewrite this paragraph for a corporate HR audience — people who work inside large organizations, not business owners. They're used to more formal language and they're skeptical of anything that sounds like startup advice.

Keep my underlying argument exactly the same. Keep it to roughly the same length. But adjust the tone, vocabulary, and framing so it lands with that audience without sounding like a different person wrote it.

What good output looks like

The rewrite adjusts register without gutting the argument. Corporate vocabulary comes in where appropriate — "workforce enablement" instead of "how you hire," "cross-functional alignment" instead of "getting everyone on the same page." But the core logic stays intact.

What should NOT happen: the rewrite becomes bureaucratic filler. It shouldn't be longer. It shouldn't lose the point. It shouldn't sound like a press release.

The tell is whether the argument survives the translation. Paste both versions side by side. If you can read the second version and still understand exactly what the first one was arguing, the rewrite worked.

What this reveals

This is the test for tonal intelligence. Claude has read an enormous amount of text across every register — technical, casual, legal, academic, journalistic, corporate. That training shows up in rewriting.

The more specific your instructions, the better the output. "Adjust the tone" is vague. "For a corporate HR audience who are skeptical of startup advice" gives Claude something to calibrate against. Try this with a piece of your own writing. The quality of the tonal shift is a good proxy for how well the tool actually reads your intent.

Prompt 4: Brief-Style Prompting

What it tests: The difference between a weak prompt and a strong brief — and what prompt quality does to output quality

The weak prompt (run this first):

Write a LinkedIn post about time management.

Run it. Read what you get. It will be adequate and generic.

Now run the brief:

Write a LinkedIn post for me. Here's the context:

• Audience: independent consultants with 5+ years of experience who feel busy but not productive
• Their problem: they track time poorly and undercharge as a result
• My argument: tracking time isn't about billing — it's about knowing where your work actually goes so you can cut what isn't paying off
• Tone: direct, slightly contrarian, no motivational language
• Format: short paragraphs, no bullet lists, no em-dashes, no emojis
• Length: under 200 words
• Don't start with "Are you..." or "Let's talk about..."

What good output looks like

The first prompt gets you a generic LinkedIn post. It's functional, bland, and indistinguishable from a hundred other posts on the same topic.

The second prompt gets you something with a specific argument, a specific audience, and a specific voice. The constraint list isn't bureaucratic — each item closes off a direction Claude would otherwise explore. Banning "Are you..." eliminates one of the most overused LinkedIn openers. Specifying "no motivational language" shifts the register from inspirational to practical.

The difference between the two outputs is large and immediate.

What this reveals

Prompt quality drives output quality more than people expect. Claude isn't psychic — it generates what the prompt implies. A vague prompt implies a generic response. A brief tells Claude who it's writing for, what it's arguing, what it should sound like, and what to avoid.

The brief format is learnable. It doesn't require technical knowledge — it requires the same clarity you'd give a human writer. If you couldn't brief a person clearly enough to get what you want, Claude won't fill in the gap. If you can, the output closes the distance considerably.

Prompt 5: Honest Uncertainty

What it tests: How Claude handles the edges of its knowledge — whether it confabulates or acknowledges limits

The prompt:

I'm trying to understand the current regulatory status of AI-generated content under EU copyright law. Specifically: does the EU AI Act address copyright ownership for outputs generated by AI systems? And what's the current legal consensus — if there is one — on who owns those outputs?

If there's meaningful uncertainty here, I want to know that. I'd rather you tell me what's genuinely unsettled than give me a confident answer that might be wrong.

What good output looks like

Claude does two things well here. First, it tells you what it actually knows — the broad strokes of the EU AI Act, the provisions relevant to transparency and high-risk systems, what the act does and doesn't address about copyright ownership.

Second, it tells you where the law is genuinely unsettled. Copyright ownership for AI-generated outputs is an active area of legal debate. Different jurisdictions have reached different conclusions. The EU framework is evolving. A good response names the uncertainty explicitly, distinguishes between what the law currently says and what it doesn't cover, and tells you this is a question a lawyer would need to answer based on your specific situation.

What bad output looks like: a confident answer that smooths over the ambiguity. Claude should not tell you "under EU law, the copyright owner is X" as though that's resolved. If it does, that's a red flag.

What this reveals

This is the test for epistemic honesty. Language models can confabulate — generate plausible-sounding answers about things they don't actually know with confidence. The better models flag uncertainty rather than paper over it.

Claude has been trained to acknowledge the edges of its knowledge. It's not perfect at this, but it's better than you might expect. Legal questions, recent events, highly specific technical claims — these are the areas where the distinction between "I know this" and "I'm generating something plausible" matters most.

When you ask Claude something you're not sure about, pay attention to whether it hedges appropriately. That hedging is a feature, not a weakness.

What These Five Prompts Tell You

Taken together, they map the territory.

Context retention means you can work with real documents, not just toy examples. Nuanced reasoning means you can use Claude for decisions that don't have clean answers. Tonal control means the output can actually sound like something you'd send. Prompt quality means your investment in learning to brief well pays off immediately. Honest uncertainty means you can trust the output more when Claude is confident, because it tells you when it isn't.

None of this is theoretical. Run each prompt. Evaluate what you get. That evaluation is faster and more reliable than anything I could tell you.

If you only have five minutes right now, start with Prompt 1. Paste any document you've written recently — a proposal, a brief, a report. Ask Claude to find inconsistencies and logical gaps only. Don't summarize. See what it catches.

These five prompts show what Claude can do. They don't cover how to chain prompts together for a multi-document project without losing context, how to handle sessions where Claude starts drifting from your original brief after the first hour, or what Claude genuinely fails at — the error patterns that catch people off guard after they've decided it works.

Related reading

• How to write better with Claude without sounding like a robot
• The things Claude is quietly better at than ChatGPT

Mark Reeves is a pen name. AI Field Guide publishes role-specific, practical guides for using AI tools in real work.