The short version
- Anthropic calls examples one of the most reliable ways to steer Claude's tone and structure.
- Be specific about format and limits, and say what to do, not only what to avoid.
- Turn real writing samples into rules a bot can follow, then test a draft.
- In a Tested setup, the first test draft had 3 misses, which became new rules.
Show real samples, write clear rules, and read every draft before it goes out.
The longer version
Anthropic's prompting guide suggests thinking of the AI as a brilliant but new employee who does not know your norms. The more precisely you explain what you want, the better the result, so be specific about the format and limits you want.
Examples do much of that work. The same guide calls examples one of the most reliable ways to steer tone, format and structure. It suggests 3 to 5 that match the real task and vary enough that the AI does not copy one pattern.
Say what you want, not only what to avoid. Instead of a rule that bans something, describe the writing you do want, such as smooth paragraphs of plain prose. Giving the reason behind a rule also helps, because Claude can generalize from it.
Save the result where it applies every time. In Claude, account instructions apply to all your conversations, and a skill can apply writing patterns based on your own writing.
A Tested setup on this site tried this on Claude. Twelve made-up writing samples became a voice profile written as rules, each backed by a line from the samples. A test email then read close to the samples but had 3 misses, such as one sentence running longer than the profile allowed. Those misses became new rules.
The rules in that profile also cut filler: no warm-up openers, and no third item added to a list just for rhythm.
Two cautions from that run. The first profile came out at 3,737 characters with a quote behind every rule, and cutting the quotes brought it under 1,500. And a test in the same chat as the samples can copy from them, so a new chat with only the profile is a fairer test. Nothing was sent: every draft still needs a human read before it goes out.
Sources
Every line on this page comes from one of these. Checked Nov 5, 2026.
Related setups
Setups on this site that put this page to work.
Make AI write in your voice
Paste 10 to 20 pieces of your own writing, get a one-page voice profile of habits and banned phrases, and save it where each bot reads standing…
For: Anyone whose AI drafts come back stiff and generic: owners writing customer emails,…
How it went: 12 made-up writing samples gave a 6-part voice profile, which was cut to 1,498 characters after a test draft showed 3 misses
ChatGPT · Claude · Grok Bot
Turn a messy draft into a clear, polite work email
Paste a rough, angry, or rambling draft and get a clear, polite, concise work email (or Slack) draft. Facts kept, no new commitments invented, you…
For: Anyone who types a vent in Notes first, then stalls before sending a professional reply…
ChatGPT · Claude · Gemini · Grok Bot
About 5–15 minutes per sticky email
Draft the Hard Emails: Saying No and Following Up Twice
Drafts polite declines, no-meeting replies, and second follow-ups
For: A professional who dreads saying no to meetings and lets follow-ups die from awkwardness
ChatGPT · Claude
Estimate: about 10 minutes per difficult email
Common questions
How many writing samples should an AI see?
Anthropic's prompting guide suggests 3 to 5 examples that match the real task and vary enough that the AI does not copy one pattern.
Where should a voice profile be saved?
Somewhere it applies every time. In Claude, account instructions apply to all conversations, and a skill can apply patterns based on your own writing.
Is one test draft enough?
Not always. In a Tested setup, the first test draft had 3 misses that became new rules, and a new chat with only the profile makes a fairer test.
New explainers and setups by email, with sources.
Newsletter
Get the free Portable AI Office, plus occasional new setups
The free Portable AI Office opens as soon as you sign up: missed-call texts, quote follow-ups and late payment reminders, drafted in the AI you already use for your approval. Plus occasional new setups by email. Unsubscribe anytime.