Prompts to Avoid: AI Prompt Patterns That Fail

Picture this: you're a marketing manager, and you've just asked AI to help you write a product launch email. The output is technically fine -- grammatically correct, covers the features, hits the right length. But it's also perfectly average. It could have been written by any AI, for any product, for any company.

That's what ai prompt patterns to avoid look like in practice. Not catastrophic failures -- just outputs that cost you time to fix and trust to ship.

The Problem a Marketing Team Faced

A demand generation team at a 300-person SaaS company had been using AI for email copy for six months. Open rates on AI-assisted campaigns were actually lower than their pre-AI benchmarks -- 18% vs. 23%. The content wasn't bad. It just wasn't theirs.

After a two-week audit, they identified five patterns in their prompts that were consistently producing generic output. None of them were obvious mistakes. All of them are incredibly common.

The Wrong Approach

Here are the five prompt patterns the team was using -- and why each one predictably fails:

Pattern 1: The Topic Dump

Write an email about our new project management feature for marketing teams.

What this does: Gives the model a topic without a task. "About" is not a task. The model fills the vacuum with the most statistically average content it has seen on the topic -- which is generic SaaS marketing copy from the last five years.

Pattern 2: The Vague Persona

Write this for a busy marketing manager.

What this does: "Busy" is not a useful persona constraint. Every persona is busy. This adds words to the prompt without adding information the model can use.

Pattern 3: The Wishlist Instruction

Make it engaging, persuasive, and professional.

What this does: Nothing. These adjectives are the model's default mode. You're asking it to do what it already does. Adjective-only tone instructions without examples or examples of the opposite are essentially noise.

Pattern 4: The Length-Without-Structure Request

Write a 300-word email.

What this does: Specifies quantity without specifying quality or structure. The model will hit 300 words by padding whatever structure it chose by default -- usually feature-benefit-CTA, which is the lowest common denominator of marketing email structure.

Pattern 5: The Open-Ended Brainstorm

Give me some ideas for subject lines.

⚠️ Common mistake: Asking for "some ideas" without specifying count, diversity requirements, or evaluation criteria. You'll get five safe, similar subject lines that all use the same structural hook. Constraint-free brainstorming produces constrained outputs.

The Correct Prompt

Here's the revised approach the team landed on:

You are a B2B email copywriter who specializes in feature launch campaigns for project management tools.

Task: Write a launch email for our new Timeline View feature, targeted at marketing operations managers at companies with 50-500 employees who currently use spreadsheets for campaign planning.

The email should accomplish one thing: get the reader to start a 14-day trial. Not to learn more. Not to request a demo. To start a trial today.

Structure:
- Subject line (3 variations: curiosity-gap / specific-benefit / social-proof)
- Preview text (matches each subject)
- Opening: one sentence that names the specific pain (spreadsheet chaos during campaign season)
- Body: two short paragraphs max
- CTA: single action, one sentence

Constraints: No adjectives that apply to all SaaS tools (innovative, powerful, easy-to-use). Feature name must appear in the first sentence. Reading time under 45 seconds.

What this does: Converts every vague element into a specific one. The task has a verb (write), object (launch email), and success criterion (trial start, not learning or demo). The persona has a specific job title and situation. The structure tells the model exactly what to produce and in what order.

Results and What Changed

Open rates on the revised campaigns: 26% -- a 44% improvement over their AI-assisted baseline and a meaningful lift over their pre-AI benchmarks.

The team also found that prompts written with this structure took about the same time to write as their old Topic Dump prompts -- the discipline just moved to the right place (upfront specification instead of downstream revision).

⚡ Pro tip: Every time you find yourself heavily editing an AI output, stop and audit the prompt before rewriting the content. 80% of the time, the problem is in the prompt, not the output.

How to Apply This to Your Situation

The five anti-patterns above aren't marketing-specific. They appear in coding prompts ("write a function to handle authentication"), research prompts ("summarize the competitive field"), and data prompts ("analyze this dataset and find insights").

In every case, the fix is the same: replace topic with task, replace adjectives with examples, replace length with structure, and replace "some ideas" with "N ideas, each with [specific format and constraint]."

⚡ Pro tip: Before sending any prompt, read it back and ask: "Could this instruction mean five different things to five different people?" If yes, it will mean five different things to the model too.

Next Steps

Audit your most-used prompts against the five anti-patterns above. It takes 10 minutes and will surface problems you've been blaming on the AI.

Once you've cleaned up a prompt, save the improved version in PromptABCD. The goal isn't to write better prompts once -- it's to stop rewriting the same bad prompt every time you start a new campaign.

Five More Anti-Patterns Worth Knowing

Beyond the core five the marketing team identified, these additional patterns fail consistently across professional contexts:

Anti-Pattern 6: The Assumed Context

Continue the project we discussed.

The model has no memory between sessions. Any prompt that assumes prior context will produce a confused or generic response. Always include the relevant context explicitly, no matter how obvious it feels.

Anti-Pattern 7: The Multi-Task Dump

Write a blog post, create 5 social captions, draft an email, and suggest 3 CTAs for our new feature launch.

What this does: Forces the model to context-switch four times in one output, usually producing mediocre versions of all four tasks. Separate tasks get separate prompts. A prompt is not a to-do list.

Anti-Pattern 8: The Recursive Instruction

Make it better.

"Better" relative to what standard? What dimension? Better at what? This instruction gives the model no vector for improvement, so it makes arbitrary changes -- sometimes improving one thing while degrading another.

⚡ Pro tip: Replace "make it better" with a specific target. "Make the opening sentence more specific -- it should name the exact pain point, not a category of pain." Improvement instructions that name the dimension of improvement produce 3x more useful edits.

Anti-Pattern 9: The Jargon-Only Brief

Write copy for a full-funnel demand gen campaign targeting ICP accounts in the ENT segment with a 3-4 week sales cycle.

Unless you're certain the model knows your specific industry's jargon, define it. "ICP" and "ENT segment" and "demand gen" all have multiple possible interpretations across industries. Undefined jargon leads to outputs calibrated to the wrong interpretation.

Anti-Pattern 10: The Perfection Request

Write the perfect headline for this article.

"Perfect" is unmeasurable. The model will produce something confident-sounding that satisfies no actual criteria. Ask for "three headline variations, each optimized for a different reader motivation: curiosity, fear of missing out, and practical benefit." Concrete variation criteria produce testable outputs.

⚡ Pro tip: Every time you catch yourself using an abstract evaluative adjective (perfect, great, compelling, powerful) in a prompt, replace it with a specific criterion. "Compelling to a CFO evaluating vendor risk" is a criterion. "Compelling" alone is noise.

Diagnosing Your Own Anti-Patterns

The fastest way to find your personal prompt anti-patterns: look at your edit history. Every time you significantly revise an AI output, the problem is usually traceable to a prompt weakness.

Keep a simple log for two weeks: what did you ask for, what did you get, what did you change, and what was wrong with the prompt? After 20-30 entries, you'll see a pattern in your own failures. Most people have 2-3 personal anti-patterns they repeat constantly.

Self-audit prompt:
Here is a prompt I wrote and the output I had to revise heavily:

Prompt: [paste]
Output: [paste or describe]
What I changed in the output: [describe edits]

Identify what was weak in the original prompt that led to the output I had to fix. Be specific about which prompt element was the root cause.

What this does: Builds prompt retrospective skills -- the ability to learn from failures systematically rather than just trying random variations until something works.

⚠️ Common mistake: Blaming the model for outputs that are actually prompt failures. The model almost always did exactly what the prompt said. The question is whether what the prompt said matched what you meant.

Once you've cleaned up a prompt and confirmed it produces the output you want, save it in PromptABCD. A well-specified prompt is a reusable asset -- not just a one-time fix.

Five More Anti-Patterns Worth Knowing

Beyond the five core anti-patterns the marketing team identified, these additional failure patterns appear across almost every domain:

Anti-pattern 6: The Hallucination Invitation

What are some statistics about remote work productivity?

What this does: Asks for specific facts without providing source constraints. The model will produce plausible-sounding statistics, some of which it invented. The fix: either provide the data yourself ("here are three statistics -- analyze them") or ask for reasoning rather than facts ("what would you expect to be true about remote work productivity, and why?").

Anti-pattern 7: The Contradiction Stack

Write a detailed, comprehensive, in-depth analysis. Keep it concise and to the point. Make it thorough but easy to skim.

These instructions contradict each other. "Comprehensive" and "concise" are opposites. "Thorough" and "easy to skim" are in direct tension. The model will average them into mediocre output that partially satisfies each constraint. Pick one priority and state it clearly.

Anti-pattern 8: The Implicit Audience

Writing for "the reader" or "customers" or "our audience" gives the model no information it can use. Every piece of writing has a reader -- you need to specify who this particular reader is. Their role, their knowledge level, their primary concern, and what action you want them to take.

⚡ Pro tip: For any prompt where the output will be read by a specific type of person, add one sentence about what that person most fears getting wrong. A CFO reviewing a financial proposal fears missing a hidden liability. A developer reading technical docs fears following instructions that won't actually work. Naming that fear shapes the output more than any "tone" instruction.

Anti-pattern 9: The Revision Request Without Criteria

This isn't quite right. Can you make it better?

"Better" is undefined. The model will make changes -- but it won't know which aspect you found wrong. Always specify what's wrong: "The opening is too generic -- rewrite it with a specific scenario instead of a general statement. Keep everything else."

Anti-pattern 10: The One-Shot Complex Task

Asking a model to produce a complex deliverable (a full strategy document, a technical architecture, a multi-section report) in a single prompt almost always produces shallow output. Complex tasks need staged prompting: outline first, then section by section, then integration. The model has finite attention in a single response.

⚡ Pro tip: If a task would take a skilled human more than 20 minutes to do well, it probably needs more than one prompt. Break it into stages and review the output at each stage before continuing.

Pattern Recognition Is the Core Skill

The underlying skill behind all of this is recognizing which anti-pattern a failing prompt contains. Once you can diagnose that "this output is generic because the task was a Topic Dump" or "this output contradicts itself because I stacked conflicting constraints," you can fix the prompt in one targeted revision instead of rewriting from scratch.

Build this diagnostic habit and your average time from bad output to good output drops from 30 minutes to 5. The model didn't get smarter -- you got faster at using it.

Save your fixed, validated prompts in PromptABCD. The anti-pattern audit is also worth running against your prompt library periodically -- prompts you wrote six months ago often contain anti-patterns you've since learned to avoid.