Keep your voice: directing Claude without sounding like AI
After this, you'll understand why Claude defaults to a generic 'AI' sound, and you'll be able to use specific context, examples, and a quick edit pass to make its output read like you instead.
Before you start
Complete Set up a creative Project so Claude remembers your world first; once Claude holds your context, this lesson teaches the habit that keeps its output sounding like you, not like a generic AI.
The idea
That bland 'AI' sound everyone complains about? It is not Claude being a weak writer. It is Claude playing it safe because you have not told it who you are yet. By default it aims for the safe middle of everything it has seen, and your job is to pull it off that middle with specific context, real examples, and a final edit. Anthropic has a name for the safe-middle look in design: the AI slop aesthetic (the bland, samey output you get when nothing steers the model). The same thing happens in words. The cure is not a magic phrase. The cure is specificity and taste.
Why the default is generic is worth understanding, because it tells you how to beat it. When you ask for "a catchy tagline," Claude blends a thousand average taglines into one average answer. It is not being lazy. It is giving you the center of the road because you did not tell it which road you are on.
The first fix is context. Compare "write a tagline for my gym" with "write a tagline for my gym, which is a tiny no-mirrors strength studio for people who hate gyms, voice dry and reassuring." The first gets you "Unleash your potential." The second gets you something that could only be yours. More true detail in, less generic out.
The second fix is examples, also called few-shot. Few-shot means showing Claude three to five samples of the style you want before you ask for more. It learns far more from seeing your three best captions than from any adjective. Paste your examples, label them "this is my voice," then ask for new work in the same vein.
The third fix is the edit pass. Even good output carries small AI tells: overused words, a slightly stiff rhythm, em dashes everywhere, a closing line that over-explains. So after Claude drafts, run one more turn: "edit this to sound more like my examples, cut anything that sounds like a robot wrote it, and keep it plain." You will feel the difference immediately.
Watch for the tells so you can name them: words like "unlock", "elevate", "seamless", and "robust"; lists where a sentence would do; a peppy, eager tone that no real person uses; and endings that summarize what you already read. None of these are your voice. Cutting them is most of the work.
Generic is the default, not your destiny. Feed Claude your specifics and examples, then edit out the robot, and the output becomes yours.
Try it (10 min)
Watch out for
- Hunting for one magic instruction that removes the AI sound. There is none; specific context, real examples, and an edit pass are the actual fix.
- Describing your voice with adjectives but never showing examples. Three pasted samples teach Claude more than ten words like 'warm' and 'punchy' ever will.
- Shipping the first draft because it reads fine. 'Fine' is the generic middle; the edit pass that cuts the AI tells is what makes it sound like you.
- Leaving the tells in: words like elevate and seamless, needless lists, a peppy tone, and endings that re-explain the point. Cutting those is most of the job.
Paste this into Claude
I want to practice pulling your output away from generic "AI" writing toward my own voice. Here is a writing sample in my real voice (paste 2 to 3 things you have written and like): [PASTE YOUR SAMPLES] Here is what I need written: [DESCRIBE THE PIECE, e.g. "a short Instagram caption announcing a weekend sale"]. Please: 1. First, write it the generic way, the safe middle, so I can see what "AI default" sounds like. 2. Then write it again, this time matched to my pasted samples: my rhythm, my word choices, my edges. 3. Point out 3 specific things you changed between the two versions, so I learn the tells to cut on my own.
What a good response looks like
Here's the practice, for a weekend-sale caption:
**1. Generic AI default:**
"Unlock incredible savings this weekend! Don't miss our seamless sale event, where you can elevate your collection with must-have pieces. Shop now and treat yourself!"
**2. Matched to your voice (dry, warm, plain):**
"Stuff's on sale this weekend. The mugs you keep almost-buying are finally a little cheaper. Saturday and Sunday, then they go back up. Go on."
**3. Three things I changed:**
- Cut the hype words ("unlock," "seamless," "elevate") that no real person says out loud.
- Traded the eager exclamation tone for your dry, calm one.
- Dropped the "treat yourself" wrap-up; your voice trusts the reader to decide.What good looks like
- You can explain why Claude's default output sounds generic (it aims for the safe middle of everything)
- You can name the three fixes: specific context, examples (few-shot), and a final edit pass
- You can point to at least three AI tells to cut (overused words, needless lists, peppy tone, over-explaining endings)
- You produced one piece that sounds like you, not like a generic AI default
When this breaks
- Breaks when you give Claude no specifics, then blame it for sounding generic. With nothing to aim at, it returns the safe average of everything it has seen.
- Breaks when you never edit. Even a good draft carries small robotic tells, and skipping the final pass leaves your voice diluted by them.
AI can help with this
Use Claude, ChatGPT, Gemini, Canva, Adobe Express, Firefly, Midjourney, Codex, or another approved creative tool based on the job. Give the assistant real brand context, examples, exclusions, and a review standard, then use human taste to choose what survives.
You can now
You can complete the lesson outcome against a real creative job, brand, asset, or campaign.
- You can explain why the creative assistant's default output sounds generic (it aims for the safe middle of everything).
- You can name the three fixes: specific context, examples (few-shot), and a final edit pass.
- You can point to at least three AI tells to cut (overused words, needless lists, peppy tone, over-explaining endings).
- You produced one piece that sounds like you, not like a generic AI default.
Key takeaways
Claude sounds generic by default because it aims for the safe middle. Feed it your specific context and real examples, then edit out the AI tells, and the output reads like you instead.
- Claude's default output is generic because it blends toward the safe average; specifics pull it off that middle.
- Anthropic calls the bland default the 'AI slop aesthetic'; the cure is specificity and your taste, not a magic phrase.
- Few-shot means showing Claude three to five real examples of your voice before asking for more in the same style.
- Always run a final edit pass: cut overused words, needless lists, peppy tone, and endings that over-explain.
- The AI tells (unlock, elevate, seamless, robust, stiff rhythm) are not your voice; cutting them is most of the work.
Go deeper