How to Stop Claude From Sounding Like AI (My Claude Project Instructions)

Every time I opened a new chat in Claude, I used to waste the first ten minutes fixing the exact same problem. I would explain what I do, tell it who I was writing for, and ask for a normal, conversational tone.

Five seconds later, it still handed me stiff paragraphs packed with words nobody says in real life, like "delve," "tapestry," "game-changer," and "in today's fast-paced world." Then I spent twenty minutes rewriting the whole draft by hand.

If you want to make Claude write like a human on the first pass, better one-line prompts are not the answer. Putting two short files inside your Claude Project instructions (or your Claude Cowork folder) fixes it permanently.

Why Claude sounds stiff out of the box

By default, Claude knows nothing about your work, your customers, or how you talk. When it lacks specifics, it fills the gap with safe corporate filler. Giving it persistent context and a hard list of banned patterns stops that habit before it starts.

1. My Claude Project memory file

Instead of re-introducing yourself in every new chat, create one plain text file called 00_MY_CONTEXT_MEMORY.md and upload it to your Project knowledge. Keep it to six short paragraphs:

• What you do and your real background in one plain sentence.
• Who buys from you (or reads your work) and what problem has them stuck right now.
• Your offers or services, their prices, and the concrete result they deliver.
• Real numbers, past wins, or experience you can prove (plus a strict rule never to invent fake stats).
• Your main business goal for the next 90 days.
• Your working ground rules: "Write in plain English. Skip flattery. Tell me straight when my idea has a hole in it."

If you hate staring at a blank page, paste those six points into Claude and ask it to interview you three questions at a time, then write the finished file for you. It takes four minutes once, and you never have to repeat your background again.

2. My Claude banned words list (copy and paste)

Telling an AI model to "sound natural" rarely works because "natural" means different things to different people. Giving your Claude custom instructions a strict blocklist works right away. Drop this block into a second file called 01_HUMAN_VOICE_RULES.md inside your Project:

Plain-English Voice Rules: Write at a clear Grade 8 reading level. Keep paragraphs between 1 and 3 sentences. State the main point in the very first sentence with zero throat-clearing. Do not use long em-dashes.

Never use these words: delve, unleash, unlock, elevate, supercharge, harness, foster, leverage, streamline, tapestry, landscape, testament, game-changer, cutting-edge, seamless, robust, pivotal, realm, synergy, holistic.

Never use the "not just X, it's Y" sentence pattern (for example: "This isn't just a checklist, it's a system"). State plainly what something is and what it does.

Turning repeat work into reusable Claude skills

Once those two files sit inside your Project, drafts immediately sound cleaner and closer to your real voice. From there, anytime you catch yourself repeating the same job three or four times a week (writing client proposals, cold emails, landing pages, or weekly plans), save the exact steps as a reusable SKILL.md workflow file right beside them.

Set up just the memory sheet and the banned-word block above today, then test them on your very next email or post. You will notice the difference on the first reply.

P.S. If you want my complete ready-to-use setup without building it from scratch, I put both full Core Memory and Human Voice files, 20 installable Claude SKILL.md workflows for everyday writing, client work, sales, and operations, plus a 1-click offline prompt dashboard (along with my free 48-Hour Launch Checklist) right here: livefree1.gumroad.com