How to Stop AI from Writing Like AI

AI writes faster than you. This creates a problem.

Without constraints, AI produces generic content. Hedged language. Bullet points for everything. Filler phrases preceding statements nobody needed. Content applicable to any company, any situation, anyone.

You recognize it immediately. LinkedIn posts with identical cadence. Documentation explaining what something is but never why or what breaks. Blog posts padded with words to hit a count nobody requested.

The tool works. The default settings do not.

This article explains how to configure AI tools to produce output matching your standards instead of internet averages.

What AI Slop Looks Like

AI-generated content follows recognizable patterns. Once you see them, you spot them everywhere.

Opening patterns. The response starts with "Great question!" or "Absolutely!" or "I'd be happy to help with that!" Nobody asked a question. Nobody needed enthusiasm confirmed. These openers exist because training data contained customer service transcripts optimizing for perceived friendliness.

Formatting patterns. Every response gets bullet points, headers, and numbered lists. Three paragraphs of prose become twelve formatted items. A simple answer becomes a structured document. The AI assumes you want to skim. It formats for skimmability whether appropriate or not.

Hedging patterns. Sentences avoid commitment. "It's important to note..." precedes statements that are not important. "This varies depending on your specific situation" applies to everything and helps no one. "There are several factors to consider..." introduces factors the reader already knows. The AI hedges because hedging is statistically safe. Committing to positions risks being wrong.

Vocabulary patterns. Certain words appear constantly. "Leverage" when the writer means use. "Utilize" when the writer means use. "Facilitate" when the writer means help. "Streamline" when the writer means simplify. These words sound professional. They add syllables without adding meaning.

Structural patterns. The AI lists problems, then lists benefits restating the same information inverted. It summarizes what it said, then summarizes the summary. It adds "In conclusion" before conclusions that were already obvious. Content gets repeated in different words to appear thorough.

Content patterns. Claims arrive without sources. Recommendations arrive without tradeoffs. Examples stay generic enough to apply to everyone, meaning they help no one. "This helps organizations improve efficiency" tells you nothing. Which organizations? Improve from what baseline? By how much?

These patterns exist because AI optimizes for appearing helpful rather than being helpful. Without information about your specific needs, it defaults to patterns statistically common across its training data.

Why Defaults Produce Mediocre Output

AI models train on internet-scale text. The training data includes everything: academic papers, Reddit comments, corporate blogs, marketing copy, technical documentation, forum answers, news articles, creative writing.

When you average all writing on the internet, you get average writing.

The model learns what "helpful response" looks like by observing millions of examples. Those examples include:

Corporate blog posts written by committee and approved by legal. Every claim hedged. Every statement qualified. Nothing specific enough to verify or dispute.

Documentation written to satisfy compliance requirements rather than help readers. Thorough in coverage, useless in practice.

Marketing copy stripped of specifics by review processes. "Industry-leading solutions" because naming actual capabilities creates accountability.

Forum answers written to seem authoritative without committing to positions. Long explanations covering every possibility because the answerer does not know which applies.

Customer service transcripts optimizing for satisfaction scores. Friendly openings. Empathetic acknowledgments. Promises to help. Padding around actual answers.

The model synthesizes these patterns into default behavior. When you provide a prompt without context about your standards, audience, or preferences, the model fills gaps with these defaults.

The result is content that seems helpful. Seems thorough. Seems professional. But says little, commits to nothing, and applies to no one specifically.

The Solution: Context and Constraints

AI output quality scales with context. The more information the model has about who you are, who will read the output, and what good looks like for this specific task, the better the results.

You provide this context through configuration and prompting. Most AI tools offer multiple layers where you set expectations. Each layer serves a different purpose.

Layer 1: Persistent Instructions

Most AI tools allow you to set instructions that persist across all conversations. These run automatically. You configure them once.

Names vary by tool:

• Claude: Custom Instructions (Settings → Profile)
• ChatGPT: Custom Instructions (Settings)
• Copilot: Preferences
• Other tools: System prompts, user preferences, profile settings

What belongs in persistent instructions:

Your background and expertise level. This calibrates explanation depth. A senior engineer needs different responses than a junior employee learning the basics.

Senior systems administrator. 15 years experience with Linux, 
networking, and identity management. Do not explain fundamentals.
Assume I know how to read logs, write scripts, and debug systems.

Your output preferences. This eliminates formatting slop.

Be direct. Skip preamble and unnecessary caveats. 
No bullet points unless I request them.
Write in prose paragraphs. Use formatting only when it aids clarity.
No emoji. No enthusiasm openers.

Your domain context. This prevents generic advice.

I work in enterprise software deployments for mid-market companies. 
Time and budget constraints matter. Account for them in recommendations.

Words and phrases to avoid. This kills language slop.

Never use: leverage, utilize, facilitate, dive deep, 
it's important to note, there are several factors.
Say "use" instead of "leverage." Say "help" instead of "facilitate."

How you want ambiguity handled.

If my prompt is ambiguous, ask clarifying questions before 
answering. Do not guess at my intent.

These instructions run before every conversation. The AI knows your baseline expectations before you type anything.

Layer 2: Styles or Presets

Some tools let you define multiple output styles and switch between them. A style controls tone, vocabulary, and formatting rules for specific content types.

Names vary by tool:

• Claude: Styles (Settings → Styles)
• ChatGPT: Custom GPTs with different configurations
• Other tools: Presets, templates, personas

Why multiple styles matter:

Technical documentation needs different rules than sales emails. A troubleshooting guide needs different structure than a blog post. A Slack message needs different tone than a formal report.

A style definition includes explicit rules:

TONE: Direct and informative. Short sentences. Active voice. 
Address the reader with "you" and "your."

AVOID: clichés, rhetorical questions, unnecessary adjectives, 
setup phrases like "in conclusion" or "moreover"

BANNED WORDS: leverage, utilize, dive deep, game-changer, 
unlock, revolutionary, cutting-edge, seamless, robust

CLAIMS: Do not use unsourced statistics. Either cite the source 
or qualify as estimate based on experience.

STRUCTURE: One list per concept. Each item adds new information 
or gets deleted. No mirrored problem/benefit lists.

ENDINGS: Specific action items. No vague closers.

You select the appropriate style for each task. Technical writing gets one style. Marketing copy gets another. Internal communications get a third.

Layer 3: Conversation Context

Information specific to one task goes in the conversation itself.

Audience specification:

Write for IT directors who manage Windows environments and are 
evaluating Mac deployment for the first time. They understand 
enterprise IT concepts but do not know Apple-specific tooling.

Format specification:

Write this as a 500-word blog post. Technical audience. 
No introduction explaining what the product category is. 
Start with the problem and move directly to the solution.

Constraint specification:

Keep this under 800 words. The reader will skim. 
Front-load actionable information. Save background for the end.

Voice specification:

Write this in a direct, confident tone. No hedging. 
The reader should feel urgency to act.

Specifying voice bypasses default corporate tone. The model commits to a style instead of defaulting to committee-speak.

Source material:

Upload documents, paste text, or provide URLs the AI should reference. The more specific information available, the more specific the output becomes.

Rules That Eliminate Slop

Through trial and error, certain rules produce the largest quality improvements.

Ban Filler Phrases Explicitly

List phrases you never want to see. Add them to your persistent instructions or style definitions.

Phrases adding nothing:

• "It's important to note..."
• "It's worth mentioning..."
• "There are several factors to consider..."
• "Let's dive in..."
• "In today's fast-paced world..."
• "When it comes to..."
• "At the end of the day..."
• "In conclusion..."
• "To summarize..."

When you ban these explicitly, the AI finds other ways to express ideas. Those other ways are usually more direct.

Ban Overused Words

List vocabulary you want eliminated. The AI will use simpler alternatives.

Words to ban and their replacements:

The banned word sounds more impressive. The replacement communicates more clearly. Readers process simple words faster. They trust writers who use them.

Require Sources for Claims

Unsourced statistics read as fabrication. "Studies show that 73% of organizations..." triggers skepticism in experienced readers. Which studies? Conducted by whom? When? What was the methodology?

Rules for claims:

If you have data, cite the source. "According to Gartner's 2024 survey of 500 IT leaders..."

If the data comes from your experience, say so. "In deployments over the past three years, we measured..."

If you do not have data, qualify the claim. "Based on conversations with peers in the industry..." or "Anecdotally..."

If you have no basis for a claim, omit it entirely.

Round numbers signal fabrication. "50% improvement" sounds invented. "47% improvement based on measurements across 12 deployments" sounds real.

Prohibit Mirrored Lists

AI loves symmetry. It lists problems, then lists benefits that are the same information inverted.

Problems eliminated:

• Manual data entry
• Inconsistent formatting
• Delayed reporting

Benefits gained:

• Automated data entry
• Consistent formatting
• Real-time reporting

This pattern wastes the reader's time. They read the same information twice in different words.

Rule: One list per concept. If you state what something eliminates, do not follow with what it provides. Pick one framing.

Demand Concrete Examples

Generic examples help no one. "A company improved their processes" tells you nothing.

Concrete examples include specifics:

• Industry: "A 200-person healthcare company..."
• Size: "...managing 1,400 endpoints..."
• Baseline: "...reduced provisioning time from 4 hours per device..."
• Outcome: "...to 20 minutes per device."
• Timeframe: "...within 6 weeks of deployment."

Specifics create credibility. They demonstrate that you have done the work. Generic claims demonstrate that you have not.

If you do not have specific examples, say so: "This has not been deployed at scale yet, but based on testing..."

Specify Voice When Needed

For persuasive writing, name a persona or voice. This bypasses the default hedge.

Examples:

"Write this like a founder explaining the product to another founder. Direct. Assumes business competence. No hand-holding."

"Write this as a direct sales letter. Confident. Creates urgency. Addresses objections head-on."

"Write this like a senior engineer documenting a system for their replacement. Complete, precise, no filler."

When you specify a voice, the AI commits to it. The output has consistency and personality instead of corporate neutrality.

Require Specific Endings

Vague endings waste the conclusion:

• "Reach out to learn more"
• "Consider whether this applies to your situation"
• "There are many options to explore"
• "The right choice depends on your needs"

Specific endings tell the reader exactly what to do:

• "Clone the repository and run docker-compose up to test locally"
• "Open Settings, navigate to Profile, and add your custom instructions"
• "Email this article to your IT director with a request to pilot the tool"
• "Block 30 minutes tomorrow to configure your first style definition"

If you do not know what action the reader should take, your content lacks a clear purpose. Figure out the action before writing.

Building Your Configuration

Start small. Add complexity as you identify problems.

Week 1: Basic persistent instructions

Write three to five sentences about your background and preferences. Include your expertise level and one or two output preferences.

Test across several conversations. Note when the AI ignores your instructions or when outputs still contain patterns you dislike.

Week 2: Add banned words and phrases

Review outputs from week one. Identify vocabulary and phrases appearing repeatedly that add no value. Add them to your banned list.

This list will grow. Keep adding to it when you notice patterns.

Week 3: Create your first style

Pick one content type you produce regularly. Technical documentation, blog posts, internal emails, or proposals.

Write explicit rules for that content type. Tone, structure, vocabulary, and format.

Test the style on real tasks. Refine based on results.

Ongoing: Iterate based on output

Every time AI output requires editing, ask why. Was context missing? Was a rule needed but not specified? Did the AI ignore an existing rule?

Add rules to prevent problems from recurring. Remove rules that create new problems. Your configuration evolves as you learn what works for your writing.

The Quality Check

Before publishing AI-assisted content, run this check:

Read for filler. Every sentence should add information. If a sentence could be deleted without losing meaning, delete it.

Check claims. Every factual claim has a source, comes from stated experience, or is qualified as an estimate. No unsourced statistics.

Verify specifics. At least one concrete example with industry, size, baseline, or outcome. Nothing generic enough that a competitor could copy it verbatim.

Check structure. No mirrored lists. No repeated summaries. Information appears once.

Verify ending. The conclusion tells the reader exactly what to do. No vague closer.

Search banned words. The AI will slip banned vocabulary through. Search your banned list. Replace violations.

This check takes five minutes. It catches slop that survives even good configuration.

Why This Matters

Generic content performs worse than no content.

When you publish slop, you train your audience to skim or ignore you. You dilute your expertise into noise indistinguishable from every other AI-assisted post in their feed. You lose the trust that comes from demonstrating specific knowledge.

The writers and companies winning with AI are not the ones producing the most content. They are the ones producing content their audience cannot get elsewhere. Specific knowledge. Clear opinions. Concrete examples. Distinctive voice.

AI is a tool. You control the tool. The output quality depends on the constraints and context you provide.

Do This Now

1. Open your AI tool's settings. Find custom instructions, preferences, or profile settings.
2. Write two sentences about your background. Include your expertise level and domain.
3. Add three phrases to a banned list. Start with "It's important to note," "Let's dive in," and "There are several factors."
4. Add one output preference. Example: "No bullet points unless I request them."
5. Save your configuration. Start a new conversation. Test with a real task.
6. Note what still needs fixing. Add rules. Test again.

Your configuration will improve over time. Start now. The first version does not need to be complete. It needs to be better than no configuration at all.

Published by MacJediWizard Consulting