Claude Custom Instructions: How to Stop AI From Sounding Like AI

---

If you’ve ever pasted an AI-written blog post, email, or landing page into a doc and felt something was off — the same handful of phrases, that rhythm of short sentences broken up by dashes — you’ve met “AI slop.” It’s not your imagination, and it’s not unique to one tool. Ask it in ChatGPT, ask it in Claude, ask it anywhere: the default voice of every general-purpose AI sounds like every other business writing in the same generic tone.

The fix isn’t a better prompt. It’s Claude custom instructions — a project-level memory that permanently corrects the AI’s tone, word choices, and formatting habits so that everything it writes for you, from a single social post to a full landing page, comes back sounding like you instead of sounding like AI. This isn’t a settings toggle you flip once. It’s a short, deliberate process of taking what the AI has already written badly, correcting it by hand, and feeding those corrections back in as memory. Here’s exactly how that works.

Why Claude’s Custom Instructions Fix the “AI Writing Style” Problem

“AI writing style” isn’t one mysterious thing — it’s a pattern you can actually name. It’s the same opening lines, the same handful of transition phrases, and a rhythm of sentences broken up by dashes for emphasis. Once you’ve seen it a few times, you can spot it in seconds, and so can your readers — which is exactly why “AI slop” has become shorthand for content nobody trusts.

The reason this happens by default is that a general-purpose AI has no information about how you specifically write. It’s trained on an enormous average of internet writing, so unless you tell it otherwise, its output regresses to that average every time. That’s true whether you’re using Claude or another model — the tools are converging on the same underlying problem, which is exactly why “how do I stop AI from sounding like AI” is one of the most common things people are now typing into Google.

The fix is not a smarter prompt typed fresh each time. It’s giving the AI a standing, permanent correction — a document it checks before it writes anything, every single time. That’s what Claude’s custom instructions and project memory are built for, and it’s the difference between an AI that writes at your brand and one that writes as your brand.

Does Claude Have Custom Instructions? Where to Find Them

Yes — and if you’re using Claude through the browser only, you’re missing the feature that makes this whole process work. Custom instructions and project memory live inside Claude Desktop, under Projects, not in a one-off chat window.

Here’s where to find it:

1. Download Claude Desktop if you haven’t already.
2. Open it, go to Projects, and start a new project.
3. Name the project and write a short description of what you’re trying to achieve with it — this matters more than it sounds like it should, because it’s the first thing that orients every chat you start inside it.
4. Start a new chat from inside that project (not a general chat — this is the detail almost everyone misses, and it’s the whole reason the custom voice sometimes “doesn’t work”).

You can run one project for your whole business, or split things into a few project-specific ones — a common setup is one large project for the bulk of the business and a couple of smaller side projects for specific initiatives. Either way, the project is the container. Everything you upload and everything Claude remembers is scoped to it.

How to Train Claude’s Custom Instructions on Your Own Writing Style

This is the step people skip, and it’s the one that actually does the work. You cannot ask Claude to “sound more like me” and expect a lasting fix — you have to show it, in writing, what “like me” means. Here’s the process:

1. Pull a document Claude (or another AI) has already written for you — a blog post, an email, anything with the generic tone in it.
2. Go through it manually and rewrite it in your own words. This is the effort most people aren’t willing to put in, which is exactly why most AI content still sounds like AI. Take the exact phrases you’d never say, and write down what you’d say instead.
3. Turn that into a simple reference document: “Don’t write it like this, say it like this instead,” sentence after sentence.
4. Upload that document into your project chat and say explicitly: “Please remember these phrases and prompts. Make sure to always write according to this document, and place it in memory for future chats in this project.”
5. Claude will typically turn this into its own memory file automatically — you’ll see it appear in your project’s memory panel afterward.

Do this once, properly, and every new chat you start inside that project will draw on it automatically. Skip it, or start a chat outside the project, and you’re back to the generic AI voice — because the memory doesn’t travel with you outside its project.

What to Put in Your “Do Not Say This, Say This Instead” Document

The content of this document is the whole point, so be specific rather than vague. For each entry, write the exact phrase or word you don’t want, followed immediately by what you’d say instead. Do this sentence by sentence, not as a general style guide — “more casual tone” tells Claude almost nothing useful, but “don’t write ‘in today’s fast-paced landscape,’ just say what changed” tells it exactly what to avoid and what to do instead.

The Dash Problem, and When to Keep One

The em dash deserves its own callout because it’s one of the most reliable “AI tells” in existing content. Dashes aren’t wrong on their own; they have a real function, prompting a certain reading rhythm and creating emphasis. The problem is AI tools default to using them constantly, in places a comma or a full stop would read more naturally. The fix isn’t “never use a dash” — it’s telling Claude explicitly: only keep a dash when you’re emphasizing something specific; otherwise, replace it with a comma or a period. That one instruction, added to your memory document, removes one of the most obvious AI tells from everything it writes for you afterward.

Claude Custom Instructions for Creative Writing: Teaching It Your Brand, Not Just Your Voice

Once the tone is handled, you can take the same project and teach it your actual branding — colors, fonts, and a full brand document with taglines and focus points — so it references that automatically anytime you ask it to build something visual, like an HTML landing page or a slide deck.

The process is the same pattern as training your voice, just applied to branding instead of words:

• Upload your branding details — your specific fonts and your exact color values, not vague descriptions of them.
• Give it a concrete example. A practical way to do this: find an existing design you like the direction of (a Pinterest board works fine), and tell Claude directly — these are the colors, this is the font, this is the branding style I want.
• Ask it to compile everything into a formal brand document, including taglines and focus points, and to save that into the project’s memory the same way you did with the tone document.

Once that’s in place, the project can produce on-brand output — in your colors, your fonts, your tone — without you re-explaining any of it in every new prompt.

Claude Project Knowledge vs. Custom Instructions vs. Skills

This is a distinction that trips people up, so it’s worth being precise about it: project memory and custom instructions are not the same thing as a Claude skill or plugin.

Project memory is the tone and branding correction living inside one specific project — Claude checks it automatically every time you write inside that project. A skill or plugin, on the other hand, is a defined workflow: do step A, then B, then C. Skills are genuinely useful for repeatable processes — a monthly content-ideas generator or an SEO/AEO blog-writing workflow, for example — but they are not tied to any one project’s memory.

That has a very specific practical consequence: a skill used outside your branded project will still hand you the same generic AI voice, because it has no access to the tone and brand memory that only exists inside that project. The same skill, run inside your branded project, automatically inherits your custom voice — same workflow, completely different output. If you want a repeatable workflow to also sound like you, it needs to run from inside the project that holds your memory, not as a standalone skill elsewhere.

Can AI Copy Your Writing Style? A Real Before-and-After Example

Yes — reliably, once it has the right memory to draw from. The clearest way to see the difference is to compare the exact same prompt run in two different places.

Outside the branded project, a generic prompt tends to produce something like: “In today’s fast-paced digital landscape, it’s essential to leverage AI to unlock new levels of productivity and efficiency for your business.” It’s the sentence you’ve read a thousand times, from a thousand different AI-written pages, adding nothing distinct.

Inside the branded project — with the tone document and brand memory in place — the same prompt pulls from your own real examples and past chats instead of the generic average, and comes back sounding like something you’d actually publish. That gap between the two outputs is the entire value of doing this setup work once.

Conclusion

Claude custom instructions solve the “AI writing style” problem, but not automatically — the fix is a project with memory, built from a document you write by hand once: the exact phrases you’d never say, and what you’d say instead, plus (optionally) your full branding on top. Do that inside a Claude Desktop project, always work from that project rather than a general chat, and keep skills or plugins running inside that same project if you want their output on-brand too. The result is an AI that writes posts, landing pages, and even a full funnel in your voice and your brand colors, every time — without you rewriting its output by hand afterward.

---