How to Make Claude Sound More Human (Without Losing What You're Trying to Say)

Claude's default output is identifiable. Not because it's bad — it's often well-structured and useful — but because it uses the same set of rhetorical habits regardless of what you asked for.

The hedging openers. The em-dash pairs. The tendency to close with a summary of what was just said. The formal connectives ("furthermore," "additionally") that no one uses in actual conversation.

These aren't random. They come from how Claude was trained to be helpful and appropriately cautious. And they can be removed, but removing them is a different problem from making Claude sound like you. (Anthropic's prompt engineering documentation covers the general levers; the voice-specific ones are below.)

---

Which AI-Tells Does Claude Use Most?

Before configuring anything, it helps to know which habits you're actually trying to remove. Claude's most consistent tells:

Hedging openers:

• "Worth noting that…"
• "Important to understand that…"
• "Of course, …"
• "Certainly, …"
• "That said, …" (at the start of paragraphs)

Em-dash overuse: Claude uses paired em-dash asides at a much higher rate than most human writers, often where a comma or period would serve better.

Formal connectives:

• "Furthermore, …"
• "Additionally, …"
• "In conclusion, …"
• "It should be noted that…"
• "As mentioned above, …"

Structural over-engineering: Claude often formats responses as a series of balanced points even when the question called for a direct answer. "There are three key aspects to consider. First, … Second, … Finally, …" when you asked a yes/no question.

Summary restatement: Many Claude responses end by summarizing what was just said: useful in long documents, obvious and filler-like in shorter responses.

---

How Do You Configure the AI-Tells Away?

A Claude Project with a persistent NEVER list removes most of this without per-prompt work. Here's the block that handles the main tells:

NEVER:
- Start sentences or paragraphs with: "Worth noting", "Important to understand", "Of course", "Certainly", "That said", "Furthermore", "Additionally", "In conclusion", "As mentioned"
- Use em-dash pairs as parenthetical asides. Use commas or rewrite the sentence.
- End a response by summarizing what was just said.
- Organize a response as a list of balanced points when a direct answer would serve better.
- Add qualifications that weren't requested ("While there are many approaches...", "Every situation is different, but...")

This handles the surface layer. Pair it with an ALWAYS block that establishes the register you want:

ALWAYS:
- Write in the register of the request; if it's a casual question, answer casually.
- Use contractions where they're natural (don't, can't, it's).
- Prefer shorter sentences in places where a longer sentence creates distance.
- Answer the question directly before adding context.

---

What This Doesn't Fix

The NEVER list removes AI-tells. It doesn't give Claude your voice.

After a good NEVER block, Claude output sounds like a clear, direct, competent writer. It no longer reads as obviously AI-generated. But it still doesn't sound like you specifically; it sounds like an instructed AI that's been told to write clearly.

The gap shows up in things instructions can't fully describe:

Sentence rhythm. Your sentences have a particular ratio of short to long. Claude's instructed writing has a different ratio: more uniform, less varied. You can say "vary sentence length" but you can't specify the exact pattern.

Qualification patterns. How you hedge when you're uncertain is personal: the specific phrases, the placement, the degree. "Instructions say minimize hedging" doesn't replicate the pattern; it just removes some hedges.

Transition logic. How you move between ideas (the connectives you choose, the ideas you implicitly assume the reader shares) is characteristic. Instructions can remove formal connectives but can't add your specific connective logic.

Voice in context. Your register shifts differently from Claude's trained shifts. When you're being serious vs. informal, your patterns change in specific ways that no rule set fully captures.

---

The Voice Sample Approach

The most effective approach for ongoing work: pair the NEVER/ALWAYS rules with actual writing samples.

In a Claude Project, upload 3–5 pages of your actual writing (emails, blog posts, reports, whatever represents how you actually write for this kind of task. Claude will reference those samples alongside the instructions, which is more effective than instructions alone for matching register and rhythm.

The limitation: Claude can approximate patterns from samples, but approximation degrades over longer pieces and across varied tasks. The sample gives it a reference point; it still applies that reference through its own interpretation.

For writers who need consistent voice across extended output, and where the gap between "sounds less robotic" and "sounds like you" actually matters, a Style Profile provides a more precise extraction. Not a description of your voice, but the actual rhythmic and structural patterns extracted from your writing, built to be used alongside AI tools rather than described to them. The same approach applies across all AI models; see writing profiles for every AI model for how the extraction maps to Claude vs. ChatGPT vs. Gemini specifically.

---

Practical Setup for Claude Projects

If you're starting from scratch, here's the minimum viable Claude Project configuration for human-sounding output:

You are assisting [your name] with [type of writing: emails / articles / reports].

ALWAYS:
- Match the register of the request.
- Use contractions where natural.
- Answer directly. Add context after the direct answer.
- Write in second or first person, not passive constructions.

NEVER:
- Hedging openers: "Worth noting", "Important to understand", "Of course", "Certainly", "That said" (as opener), "Furthermore", "Additionally", "In conclusion"
- Em-dash pairs as asides
- Summary restatements at the end of a response
- Balanced-point structures when a direct answer serves better
- Qualifications I didn't ask for

MY VOICE:
[Paste 2-3 sentences that describe how you actually write: your sentence length preferences, your tone, your approach to the reader. Don't describe what you want; describe what you do.]

Upload 2-3 pages of representative writing as a knowledge file. The combination of rules plus examples does more than either alone.

For a complete walkthrough of how to configure a Claude Project for your writing, see how to make Claude sound like you. For the same configuration extended to professional writing register, see how to build custom instructions that actually capture your voice.

---

Try Voice Extraction, Not Just Rules

The NEVER list removes AI-tells. The samples give Claude a reference point. But for extended writing, or any writing where voice consistency actually matters to your readers, the most reliable method is extracted patterns, not described ones.

A Writing DNA Snapshot derives the knowledge-file content from your actual writing: the rhythms, the register shifts, the sentence structures that make your prose yours. It's the input that makes both Claude Projects and Custom GPTs work as voice tools rather than just rule-followers.