Prompt Tactics

Examples Beat Adjectives: The Single Biggest Claude Prompt Upgrade

Published April 27, 2026 · 8 min read

You ask Claude for "a casual product update." It hands you back something that opens with "Hey friends!" and ends with "Stay tuned for more updates!" You delete it.

You try again. "Make it casual but smart, like Stripe's blog." The output gets a little better, but it's still got that LinkedIn-with-extra-emoji feel. You spend the next twenty minutes editing every sentence.

This isn't a model problem. It's an adjective problem.

The single biggest upgrade you can make to any Claude prompt — emails, landing pages, captions, scripts, internal updates, anything — has nothing to do with phrasing tricks or magic words. It's swapping vague adjectives for concrete examples. One paste, every time. The output difference is closer to "different writer" than "same writer with edits."

Why adjectives quietly sabotage your prompts

Words like "casual," "professional," "engaging," and "punchy" sound specific in your head. They aren't. They're ambiguity buckets. Claude has read tens of millions of pages where each of those adjectives was attached to wildly different writing styles. Its training data includes:

• "Casual" emails from luxury-brand newsletters that read like soft poetry
• "Casual" emails from SaaS teams that read like a Slack message at 3pm
• "Casual" emails from food bloggers that read like a recipe intro
• "Casual" emails from tech CEOs that read like a half-edited tweet

When you say "casual," Claude has no way to pick. So it averages. The average of every casual writing style ever written is the kind of friendly-but-corporate voice you see on every SaaS homepage made between 2018 and now. That voice is the LLM mean. It's not Claude being lazy. It's Claude being honest about the underspecified word you handed it.

Same problem with "professional." Same with "punchy." Same with "engaging." They're all averages waiting to happen.

What an example actually does to the model

An example is a thousand decisions made for Claude in advance.

When you paste a single piece of writing and say "match the energy of this," the model doesn't have to guess at sentence length, paragraph rhythm, comma density, vocabulary register, sign-off style, capitalization habits, hedging frequency, sentence-fragment tolerance, em-dash use, parenthetical asides, opening-line structure, or about forty other dimensions of voice. The example resolves them all at once.

This is why few-shot prompting outperforms zero-shot prompting in basically every academic benchmark, and why it does the same in your day-to-day Claude use. You're not "giving Claude an example." You're shrinking the space of possible outputs from millions to dozens.

And here's the part that surprises people: two examples beat one example by a wide margin. With one example, Claude can still wander toward the LLM mean if the example feels too unusual. With two, it triangulates. It infers the shared voice between the two — the part that's deliberate — and discards the noise. Two is the sweet spot. Three is fine. One is risky. Zero is the LinkedIn voice you've been fighting.

The exact upgrade, in one prompt

This is the move. Take any prompt you'd normally write with adjectives:

Write a casual product update announcing our new
Slack integration.

Replace the adjective with two examples:

Write a product update announcing our new Slack
integration.

Match the voice and structure of these two updates
(both written by us, both performed well):

---
EXAMPLE 1:
[paste your last update that landed]
---
EXAMPLE 2:
[paste another update that landed, different topic]
---

Constraint: 110-160 words. No greeting. No "we're
excited to announce." End with one sentence about
what to do next.

That's it. That's the entire upgrade. The output you get back will read like a third update in the same series rather than a new SaaS announcement written by a stranger.

You can stack adjectives on top if you want. They don't hurt. But the examples are doing 90% of the work, and on their own they'd carry the prompt without a single adjective.

Where to get examples when you don't have your own

The best examples are pieces you wrote (or your team wrote) that performed well — they tell Claude exactly how your specific audience already responds. But you don't always have those, especially when you're starting out or writing for a new format. Three fallback sources, in order of usefulness:

1. Pieces from people whose voice you wish you had

If you're a solo founder and you wish your updates sounded more like Mariah Coz's emails or Patrick McKenzie's posts, paste two of theirs. Tell Claude: "Match this voice. I'm not copying their content, I'm using their pacing and tone as a reference." This is fine ethically, fine legally (you're producing original text), and it's what professional writers do — they read voraciously and absorb rhythms.

2. Pieces from your category's best performers

If you're an agency owner writing a case study and you don't have one yet, find two case studies from agencies you respect. Paste them. Specifically tell Claude what you like about each ("Notice how they use specific numbers in the second paragraph"). This becomes a richer prompt than any adjective could ever be.

3. Pieces from outside your category that have the energy you want

This is the secret-weapon move. If you want your B2B SaaS landing page to feel less like every other B2B SaaS landing page, paste copy from a DTC brand, a solo creator's about page, or a YC company from 2010 that wrote like humans. The example doesn't have to be the same format. It just has to have the voice you're chasing.

The four mistakes that quietly cancel the example

Examples work. They also fail in predictable ways. If you paste two examples and the output still feels off, it's almost always one of these:

Mistake 1 — Examples that conflict

If example 1 is a 600-word essay and example 2 is a 40-word punchline tweet, Claude can't triangulate. It picks one and ignores the other, or splits the difference into something neither voice. Pick examples that share their core voice DNA. Different topics are fine. Wildly different formats are not.

Mistake 2 — Examples that aren't actually good

People paste examples that they think represent their voice but actually represent their voice on a bad day. Re-read your examples before you paste. If you wouldn't proudly send the example as a fresh message today, it's not the right example.

Mistake 3 — Adjectives that contradict the examples

If you paste two warm, conversational examples and then add "keep it formal and authoritative," Claude has to pick between the example and the adjective. It will usually pick the adjective and ignore the example, because instructions outrank style references in its training. Pick one or the other. If your examples are right, you don't need the adjectives.

Mistake 4 — Forgetting the explicit instruction

Pasting examples without saying "match the voice and structure of these" sometimes works. Sometimes Claude treats them as background context and writes in its default voice anyway. Always include one sentence that explicitly tells the model what to do with the examples.

Anti-examples: the bonus move

Once you've got two positive examples in the prompt, you can add one piece of writing labeled "do not write like this" and the output gets even tighter. This is the anti-example.

Write the post.

Voice: match these two (good):
[paste 1]
[paste 2]

Voice anti-example (do NOT write like this):
[paste a typical SaaS announcement that you find
bland or generic]

Notice especially: the way the anti-example uses
"thrilled," "seamless," and "we're proud to" — none
of those words should appear in your output.

The anti-example doesn't replace your positive examples. It sharpens them. Claude now has a vector to push away from as well as a target to push toward. The output ends up further from the LLM mean and closer to your actual voice than positive examples alone could deliver.

Use anti-examples sparingly — one is plenty. Two starts to confuse the prompt. But one well-chosen anti-example is worth four adjectives.

This applies to almost everything

The example move isn't just for tone. It works for every dimension you'd normally describe with abstract words:

• Structure — paste two pieces with the structure you want, instead of writing "use H2s and short paragraphs"
• Length — paste two pieces of the right length, instead of "around 800 words"
• CTA style — paste two emails whose CTAs land softly, instead of "soft CTA, no pressure"
• Code style — paste two functions written the way your codebase writes them, instead of "match our coding conventions"
• Hook style — paste two reel hooks that retained well, instead of "scroll-stopping, curiosity-gap opening"
• Caption style — paste two captions that drove saves, instead of "punchy, value-dense, hook-driven"

Every time you reach for an adjective, ask: do I have two pieces I could paste instead? If yes, paste them. If no, write more — your future prompts will be stronger because you'll have a library to pull from.

The checklist

Before you hit enter on any Claude prompt where voice, style, structure, or tone matters:

1. Identify every adjective you used to describe what you want.
2. For each one, ask "do I have two pieces of writing that already demonstrate this?"
3. Paste them with one explicit instruction: "match the voice and structure of these."
4. Optional: add one anti-example labeled "do not write like this."
5. Re-read your examples — would you proudly ship them today as fresh writing?

That's the upgrade. Two minutes of paste-and-label saves you twenty minutes of rewriting Claude's first draft.

Why this is the move that matters most

If I had to pick the single highest-leverage prompting habit — the one that produces the biggest delta in output quality across every use case I've tried — this is the one. Not chain-of-thought. Not extended thinking. Not better adjectives. Examples.

Claude is a precision instrument that rewards specificity. Adjectives are the lowest-resolution input you can give it. Examples are the highest. When you upgrade from one to the other, you stop fighting the model's averaging tendency and start using its imitation strength.

Every other prompt-engineering trick is a footnote next to this one. Get this right and most of the other things you're worrying about (hedging language, generic openings, mismatched voice, corporate filler) just stop happening.

One paste. Every time. That's the upgrade.