How to stop Claude from sounding like an AI: the 15 patterns to block

Improving AI writing style is not about writing better prompts. It is about banning. Every language model defaults to the same patterns unless you explicitly block them — 'Certainly!', bullet lists for every answer, 'comprehensive' as the default adjective for everything. Block them one by one, and Claude suddenly writes like a person.

This article gives you the complete list of 15 patterns and a ready-to-use template you can paste directly into your custom instructions today.

Why AI output is recognizable

Language models are trained on text that human raters scored as 'helpful' and 'complete'. That has a side effect. What raters rewarded was structured text with lists, summaries, and a friendly opener. The model learned that correlation and applies it everywhere — even where lists do not fit, even where three bullet points inflate a single thought into seven.

Research from Stanford NLP Lab (2024) found that more than 85% of AI-generated text contains the same 12-20 structural patterns regardless of model or question. Readers recognize those patterns — even when they do not know they are reading AI output. They feel it in the rhythm.

The problem is not in the content but in the packaging. Claude knows what you are asking and gives the right answer. But the packaging reveals the model. A human writer does not open with 'Certainly! I would be happy to help you with this request.' They just answer. Once a reader identifies AI output, the text loses authority — even if the content is completely correct.

There are two types of recognizable AI patterns. Structural patterns: always an introduction, always a three-part structure, always a closing. Language patterns: specific words and phrases that people never use but models produce repeatedly. You need to address both.

The 15 patterns to block

Below are the 15 writing patterns that appear in almost every AI output. Some are falsely friendly. Some are structural filler. All of them reveal that a model wrote it, not a person.

1. "Certainly!" / "Certainly, I'd be happy to help" / "Absolutely!" as an opener — a reflex that no professional writer uses.
2. "As an AI language model..." disclaimers in the middle of an answer. The reader already knows Claude is an AI.
3. "In today's rapidly changing digital landscape..." as an opening paragraph. Every AI text that starts this way is instantly recognizable.
4. The three-part structure as a crutch: "First... Second... Finally..." even when the information is not a list.
5. Bullet lists for single ideas that deserve one sentence. Not everything is a list.
6. "It is important to note that..." — a sentence that rarely adds information.
7. "Comprehensive" as the default adjective: "a comprehensive approach", "a comprehensive list", "comprehensive research". The word says nothing.
8. "I hope this helps! Let me know if you have any questions." as a standard closing. No professional email ends this way.
9. Section summaries at the end of a paragraph the reader just read: "In summary, we have seen that..."
10. Alliterative headings as structural decoration without substantive reason: "Results. Risks. Reflections."
11. "...and this is just the beginning" as a closer that gives the reader nothing.
12. Pseudo-empathetic openers: "I understand this is a challenging issue for you." Skip to the answer.
13. Stretching lists from three to seven points for appearance. Five points that are really three is padding.
14. Passive constructions for responsibility: "It should be noted that..." Write who does what.
15. Overuse of "crucial", "robust", and "holistic approach". Each one a separate AI buzzword the model overproduces.

You do not need to memorize this. Write it once, put it in your custom instructions, and Claude follows it permanently.

Building an anti-style template

An anti-style template is a ban list you write once and put in your custom instructions. It works as a negative style guide: you tell Claude what you never want to see in output. The more concrete, the better. 'Do not write vaguely' does not work. 'Never use the word comprehensive as an adjective' works.

Here is the base template. Adapt it for your context, role, and organization:

WRITING STYLE — BAN LIST

Never use the following patterns:

Openers:
- "Certainly!", "I'd be happy to help", "Great question!", "Absolutely!", "Of course!"
- "In today's rapidly changing digital landscape..."
- Pseudo-empathy: "I understand this is a challenging situation for you."

Mid-text:
- "As an AI language model..."
- "It is important to note that..."
- Three-part structure: "First... Second... Finally..."
- Empty adjectives: "comprehensive", "robust", "crucial", "holistic approach"

Closings:
- "I hope this helps!"
- "Let me know if you have any questions."
- "...and this is just the beginning."
- Summaries of what the reader just read.

Structure:
- No bullet list for a single idea.
- Do not stretch lists. Three points if there are three.
- No alliterative headings as decorative filler.

Write directly. Start with the message. Stop when the message is finished.

This template blocks the most recognizable patterns. After one week of use, notice which ones Claude still produces for you — add them to the list. On average it takes two weeks before the list is stable.

Before and after examples

The same request — 'Write a short project progress update for the board' — with and without the anti-style template.

The right column uses fewer words. It also contains more information. That is the effect of an anti-style template: less padding, more content. The board reads what is happening in five seconds.

A second example: a technical recommendation without a template opens with 'Certainly! I would be happy to help you with your question about cloud storage.' With the template, the same question starts with 'Azure Blob Storage fits better than S3 if you are already in the Microsoft ecosystem. Three reasons.' The content is the same. The credibility is different.

Combining with custom instructions

An anti-style template is most powerful as part of broader custom instructions — not as a loose prompt you paste in every time. Claude remembers nothing between conversations unless you configure it. Set it once, get permanent results.

In Claude.ai, set custom instructions via Settings → Profile → Custom instructions. Add the ban template as a separate block after your style instructions. The order does not matter — Claude reads all active instructions at once.

In Claude Enterprise, you can set the ban list per project via project instructions. That has an added benefit: every team member working on that project automatically uses the same style rules — without each person needing to update their personal instructions. One admin updates the project instructions, ten colleagues benefit.

Practical maintenance: review your template every quarter. AI models are updated and new patterns appear. What was recognizable AI output in 2025 is not always the same in 2026. Keep the list current. An outdated template is better than no template, but an updated one is always better.

Testing whether the template works

Use these three prompts to test whether your template is effective. Prompt one: 'Write an introductory paragraph for a report on automation.' Does Claude open with 'Certainly!'? The opener ban list is not working. Prompt two: 'Give me five reasons why X.' Does Claude turn it into seven? The list-stretching block is not working. Prompt three: 'Write a summary of our conversation.' Does Claude start with 'In summary, we have seen that...'? The summary block is missing.

Every pattern that gets through, add it verbatim to the ban list. More concrete is always better: not 'no openers' but 'never start with Certainly!, Absolutely!, Great question! or similar affirmers.'

Want to roll this out organization-wide — a custom ban list per department, shared via Claude Enterprise Projects? We build the complete instruction set as part of our Claude adoption program. Describe your situation via the contact form. We respond within one business day.