When I started running four sites simultaneously, I noticed a problem: every Claude-generated article was starting to sound the same.
"In recent years, X has been gaining attention..." — the same opening, repeated across Claude Lab, Gemini Lab, Antigravity Lab, and Rork Lab. Working alone on these sites, I had come to believe that the line between forgettable and memorable writing runs through voice. And I was losing mine to generic AI prose.
The root issue wasn't Claude — it was my system prompt. "Write in a warm, approachable tone" is a goal, not a description of how to write. Claude will achieve that goal using the most common patterns from its training data, which happen to be the patterns every other AI blog uses too.
This post shares what I found when I broke down voice into its actual components and started encoding them explicitly.
Why Adjectives Alone Don't Work
"Warm," "approachable," "natural-sounding" — these describe how a reader should feel, not the mechanics of how to write. Claude is very good at producing text that achieves those goals, but it'll achieve them using whatever patterns it deems most reliable.
Encoding voice requires breaking it into four concrete dimensions:
- Sentence endings and verb forms: In Japanese I specify exact endings like 「〜かもしれません」and 「〜ではないでしょうか」. In English, this means choosing between "I found that..." (first-person assertion), "It seems that..." (hedged), and "This suggests..." (analytical)
- Sentence length and rhythm: Do you write long, flowing sentences with multiple clauses, or short punchy ones? Specifying a target word count per sentence makes a surprising difference
- Point of view: First-person experiential ("I tried this and...") vs. third-person authoritative ("Research shows...") vs. first-person reflective ("Looking back, I think...")
- Emotional register: Do you state conclusions confidently, or invite the reader to think alongside you?
Once I started specifying these four dimensions separately, the voice consistency improved dramatically.
A Before/After Example
Before (system prompt said only: "Write in a clear, approachable way"):
The importance of proper system prompt design in AI applications cannot be overstated. In this article, we will explore the key principles and best practices for crafting effective system prompts that yield consistent, high-quality results.
Generic. Could have been written by anyone — or anything.
After (system prompt with explicit voice encoding):
When I started running four sites simultaneously, I noticed a problem: every Claude-generated article was starting to sound the same.
Same topic. Completely different feel. The only thing that changed was the system prompt.
Here's the specific language I added:
[Writing Voice]
Point of view: Write in first person. Use "I found," "I noticed," "In my experience" rather than passive constructions or anonymous authority ("research shows," "experts say").
Sentence length: Aim for 15-25 words per sentence. If you need a longer explanation, break it into two sentences rather than adding more clauses.
Hedging style: Prefer "I think," "it seems," "this might be" over confident declarations when discussing judgment calls. Use confident declaratives only for factual statements.
Opening rule: Never start an article with "In this post, we will..." or "In recent years, X has been..." Start with a specific situation, problem, or discovery.
Few-shot examples (write in this tone):
Example A: "I tried the official approach first, and it worked — until I added authentication. Then everything broke in a way the docs didn't cover."
Example B: "After ten years of shipping apps solo, I've stopped trying to write perfect code before shipping. The feedback from real users has always been more valuable than the extra week of polish."
Why Few-Shot Examples in System Prompts Work Best
Out of all the techniques I tried, adding two or three example sentences to the system prompt had the biggest impact.
Claude is a pattern-matching system. Concrete examples of "write like this" communicate what abstract instructions like "be conversational" cannot. The model can read the rhythm of your sentences, your vocabulary range, and your relationship to certainty — and replicate it.
Three principles for effective few-shot examples in system prompts:
-
Use your own writing: Pull from blog posts, tweets, notes — anything you've actually written. The best samples of your voice are your past sentences.
-
Keep examples consistent in tone: Don't mix one formal example with one casual one. Claude will average them, which often produces neither.
-
Keep them to 3-5 sentences each: Longer examples risk Claude getting attached to the content of the example rather than its form.
I update these examples over time. When I write something and think "that felt right," I add a few sentences from it to the system prompt. It's a slow, organic way to build a voice library.
How I Differentiate Across Four Sites
Each of my four sites has a slightly different audience:
- Claude Lab: Indie developers, solo builders, people running small projects without a team
- Gemini Lab: Google ecosystem users, people integrating AI into existing Google Workspace workflows
- Antigravity Lab: Curious generalists interested in where AI intersects with physics, creativity, and unconventional ideas
- Rork Lab: First-time app builders, people new to no-code/low-code
The base system prompt is shared across all four. The adjustments are small additions that shift emphasis:
- Claude Lab adds: "Prioritize what actually works in a solo dev environment over theoretically correct solutions"
- Gemini Lab adds: "Write from the perspective of someone already inside the Google ecosystem, not someone evaluating whether to join it"
- Rork Lab adds: "Assume the reader has never shipped an app. Every technical term needs a brief parenthetical explanation"
Sharing the base means improvements propagate across all four sites. It also means errors propagate — so I review the base system prompt monthly.
A Self-Check for Voice Consistency
After generating with Claude, I run through three quick checks:
Does the first sentence of the opening paragraph start with a specific situation or problem? If it starts with "In this article..." or "X is an important concept in...", the voice instruction isn't landing.
Does first-person appear naturally within the first two paragraphs? If "I" or "my" doesn't appear at all in the opening, the writing has drifted toward generic authority-voice.
Does the conclusion tell the reader one specific next action? If it's a bulleted recap of the article's main points, it's the wrong format. Good endings answer "what should I do next?" with a single concrete answer.
I've set up a small checking prompt that I run after generation: "Evaluate this article against these three criteria and flag any failures." Claude is usually reliable at catching its own departures from the instructions when you give it explicit criteria.
Where to Start
Pick one piece of writing you've produced that felt genuinely like you — a post, a note, an email. Find the paragraph that felt most natural to write. Put it in your system prompt under "write in this tone, using these sentences as examples."
That one change will do more for voice consistency than any combination of adjectives ever could.
Hirokawa Masaki is an artist and indie developer based in Japan, with 17 international awards and over 50 million cumulative app downloads since 2014. He runs Claude Lab and three other AI-focused sites independently.