Florian Bauer spent three weeks convinced that AI was not going to work for his team. He was head of strategy at a logistics firm in Hamburg, and every time he tried to use an AI assistant to help with competitive analysis, the outputs were vague, generic, and kind of useless. "It sounds smart," he told me, "but it never says anything I couldn't have written myself in five minutes." Then a colleague sat next to him and watched him type a prompt. "That's the problem," she said. "Your prompts are the issue, not the tool."

Florian is not unusual. Most people who struggle with AI outputs are not dealing with a capability problem; they are dealing with a prompting problem. And prompting problems cluster into recognizable patterns. You can often learn faster by studying what not to do than by studying what to do, because anti-patterns are the mistakes that feel right intuitively and fail in practice. That is exactly what makes them insidious: they seem reasonable while you are typing them. Most people repeat them for months without ever noticing the pattern. This lesson covers the most common ones, with concrete before-and-after examples.

Why Anti-Patterns Are Worth Studying

Learning to recognize anti-patterns is like learning to recognize poor design in architecture. Once you know what to look for, you see it everywhere, and once you see it, you start avoiding it automatically. Recognition is the whole skill. Underneath every pattern below sits the same gap: a bad prompt feels good when you write it, because it seems reasonable, specific, or thorough to the person who already knows what it means. Language is ambiguous, and the model has none of your mental context. What is crystal clear to you can be confusing, or can simply mean something else, to the model reading it.

Anti-Pattern 1: Vague Instructions

The most common prompting mistake is treating the AI like a mind reader. You have a clear picture in your head of what you want; the AI has access only to the words you typed. Ask it to Summarize this report. and it will produce a summary. But a summary for whom? At what length? Emphasizing which aspects? For what purpose? You get something that covers everything and is useful for nothing specific. The same failure appears at article length. Write an article about remote work. gives no guidance on audience, length, angle, perspective, or depth, so the model falls back on default assumptions that probably do not match yours.

Specificity is the fix. Summarize this report in 150 words for a VP of Finance who has not read the background. Focus on the three biggest cost drivers and what we are recommending to address each one. carries an audience, a length, a focus, and a purpose, and the output changes accordingly. The longer request works the same way: Write a 1500-word article about remote work best practices for managers who are new to leading remote teams. Focus on communication, accountability, and culture. Assume they are skeptical about remote work's viability. Include practical recommendations they can implement immediately.

Vagueness persists because what is clear in your head feels obvious, so the useful habit is an audit. Read your own prompt hunting for words that sound like instructions but are not: good, interesting, relevant, thorough, engaging. Each means something different to every reader, and the model has to guess which meaning you had in mind. Replace them with the behavior you actually want. Instead of "make it engaging," write use conversational tone, include one surprising statistic, and provide actionable tips. The test to apply: if a smart colleague would need to ask clarifying questions before starting, so does the AI.

Anti-Pattern 2: Overloaded Prompts

The opposite problem is packing too many requests into one prompt. Overloading happens for an understandable reason: you want good output, so you assume that more requirements produce better output. There is a balance point, and past it every extra constraint costs you quality rather than buying it. The model no longer knows which requirements to prioritize, so it makes those choices for you, and nothing gets the attention it deserves. What comes back is a compromise across all your goals rather than an excellent version of any one of them.

A prompt like Analyze our competitive position, identify market opportunities, write an executive summary, suggest three strategic options, evaluate the risks of each option, and format it as a presentation outline. is six tasks in a single sentence. Prompts get worse still when the requirements start contradicting each other, asking for a friendly but professional tone that also covers productivity, security, and culture, hits a word count, and carries a strong call to action. Every clause is reasonable on its own. Together they leave the model no room to do anything well.

The remedy is ruthless prioritization followed by sequencing. Identify the three most important things the output needs to have, put those in the prompt, and cut the rest. Anything you cut becomes a follow-up prompt or a refinement pass. Start with the competitive analysis; when you have reviewed it, prompt for the market opportunities; build the deliverable in stages. You keep control of quality at every step instead of receiving one large output you cannot easily repair, and it is better to get one thing right than three things mediocre.

Anti-Pattern 3: Assuming Shared Context

Each AI conversation starts fresh. The model does not know your industry, your organization, your team dynamics, your competitors, or what you discussed last Tuesday. Ask it to Suggest improvements to our product roadmap. and you are assuming it knows what your product is, what your market is, what your constraints are, where you sit competitively, and what your customers need. It knows none of that. The context that is obvious to you, because you live inside it every day, is invisible to a system that only knows generalities.

Compare that with a prompt carrying the situation on its face: We build a CRM for small e-commerce businesses with fewer than 50 employees. Our customers mainly use us for customer segmentation and email campaigns. Our main competitors are larger players. Our differentiation is ease of use for non-technical users. What are three high-impact roadmap items that would strengthen our position against these competitors while staying true to our ease-of-use promise? The same shift rescues everyday tasks. Draft talking points for the meeting with the client on Thursday. becomes usable once it names a thirty-minute renewal discussion with a mid-size retail client who signed 18 months ago, the client's concerns about support response times, the goal of retaining them at current contract value, and a tone that is direct but partnership-oriented.

Build the habit as a context audit. Before you send a prompt, ask whether someone who knows nothing about your situation could act on it. If the answer is no, add the facts an outsider would not have. Think of it as priming each conversation the way you would brief a capable new colleague: give them the who, the what, and the why before you ask them to do the work.

Anti-Pattern 4: Poor Format Specification

AI models have default output styles. Left to themselves they tend toward flowing paragraphs, moderate length, and a somewhat formal register. If that is not what you want, say so. Give me ideas for improving customer retention. leaves the format, the count, the depth, and the organization entirely open. Say instead Generate seven ideas for improving customer retention. Format each idea as: [Title]: [2-3 sentence explanation] | [Effort: Low/Medium/High] | [Expected Impact: Low/Medium/High]. Organize them by effort level, starting with low-effort ideas. and there is nothing left to guess.

Format specification matters most because of what happens downstream. If you need a table, ask for a table. If you need five bullet points, specify five. If you need a one-paragraph executive summary, say "one paragraph." If the text is going into an internal message rather than a consulting report, say that too, because pasting into a slide and pasting into an email want different shapes. The AI cannot see where the output is going, but you can tell it, and where the format is unusual, showing a short example beats describing it.

Anti-Pattern 5: No Guardrails or Constraints

Closely related, and just as common, is leaving the boundaries open. Write a summary of this article. sets no limits at all: the summary could run 100 words or 1000 words, arrive as prose or as bullets, stay at the surface or dive deep. Whatever the model picks, you will find yourself editing it back toward the thing you had in mind but never stated. Guardrails are what keep the output inside the territory you care about, and they cost one sentence to add.

Useful guardrails cover length, depth, structure, what to include, and, just as importantly, what to leave out. Write a 3-4 sentence summary that focuses on the main finding and its business implication. Do not include background information or context. Be direct. constrains all five. Exclusions carry surprising weight; telling the model what not to do prevents the wandering that produces technically correct output you cannot use.

Anti-Pattern 6: Expecting Perfect Output on the First Try

Many people send one prompt, get an imperfect result, and conclude that AI does not work for this task. That is like asking a colleague for a first draft and deciding they cannot write when the draft needs revision. The expectation usually comes from search: you type something, you get the right answer. Complex outputs rarely work that way. Writing requires revision and design requires iteration; prompting is no different, and even experienced practitioners iterate as a matter of course rather than as a sign that something went wrong.

The productive workflow is short and repeatable. Send an initial prompt to get oriented. Read the output and name precisely what is right and what is wrong. Send a follow-up that preserves what works and redirects what does not, in the manner of Good structure. The tone is too formal for this audience, so make it conversational. Also, the second section is too long; cut it to three sentences. Then repeat until the output meets your standard. The best results usually arrive after two or three passes in which you analyze the output, form a hypothesis about what is wrong, and refine. Budget time for that in your workflow instead of treating it as overrun.

Anti-Pattern 7: Inconsistent Outputs

Sometimes the same prompt produces different results each time you run it. The formatting shifts, the tone drifts, the structure reorganizes itself. The root cause is that without specific examples or constraints, models default to variation. Variation is a gift in creative writing and a disaster anywhere consistency matters, which is most professional work: customer communications, branded content, anything a team produces jointly and a reader is meant to experience as one voice.

Telling the model to Write in a professional but approachable tone. is weaker than showing it: Write in the same tone as this example: [paste example text]. This technique, called few-shot prompting, anchors output to a concrete reference, and one good example is often worth fifty words of style description. Where you need many outputs to match, specify a template that all of them must follow, as in Generate five product ideas in this format: [TITLE]: [One-sentence description]. Each idea should target a specific customer segment and solve a specific problem. Use a tone that is friendly and direct, not hype-driven.

Anti-Pattern 8: Misunderstanding What Models Can Do

Language models are exceptional at some tasks and structurally incapable of others. They do well at drafting, summarizing, reformatting, explaining, brainstorming, translating, editing, generating variations, and reasoning through problems that can be described in language. They fall short on accessing current information, since they have a knowledge cutoff; on precise numerical calculation, which they approximate rather than compute; on tasks that require real-world action, since they produce text rather than actions; and on highly specialized or proprietary knowledge that was never in their training data.

Requests that violate those fundamentals cannot be fixed by better wording. Asking for a completely novel idea that has never been thought of misunderstands a system that synthesizes from patterns in its training data. Asking about events after the knowledge cutoff asks for information the model does not have. Asking it to answer a subjective question objectively asks it to remove something intrinsic to language. The mental model worth carrying is deliberately unflattering: these are statistical pattern matchers trained on data, which synthesize from patterns, do not think, do not have true understanding, and sometimes state false information with complete confidence. When output is wrong, check whether you asked for something outside the system's capabilities before concluding it is generally unreliable.

Anti-Pattern 9: Assuming the Model Knows Your Jargon

Specialized shorthand is efficient among colleagues and lossy with a model. What should we do about our CAC:LTV ratio and our MRR churn? assumes a shared reading of customer acquisition cost, lifetime value, monthly recurring revenue, and churn rate, and assumes the model knows how your organization calculates and uses them. Some of that it can guess; the parts specific to you it cannot. Internal acronyms, product code names, and team abbreviations are worse still, since they may collide with entirely unrelated meanings the model learned elsewhere.

The fix is to describe the situation rather than to label it. Our customer acquisition cost is $500 per customer. Their lifetime value is $2000. We are losing 8% of customers monthly. What should we focus on to improve unit economics? gives the model the actual numbers and relationships instead of the shorthand. If you must use specialized terms, define them the first time you use them. Concrete facts beat compressed vocabulary in nearly every prompt.

Anti-Pattern 10: Unnecessary Complexity

Some people compensate for vagueness by writing very long, elaborate prompts: paragraphs of instruction, nested conditions, detailed scoring rubrics. That helps in certain cases and often does not, because a convoluted prompt confuses a model the same way it would confuse a person. Instructions that contradict one another, or that pile up more conditions than can be held at once, produce inconsistent output. Consider a prompt demanding specialized quantum computing vocabulary and assuming advanced knowledge of physics and mathematics while also asking for an explanation accessible to beginners. It is self-contradictory and far longer than it needs to be. Explain quantum entanglement in simple terms that a high school student could understand. Avoid technical jargon. does the job.

Edit ruthlessly. Remove redundancy, remove contradiction, remove over-specification. The goal is clarity, not comprehensiveness, and many complex prompts turn out to be verbose versions of simple ones. Ask whether you could say the same thing in fewer words without losing meaning, and if you could, do it. Start simple; if the simple prompt does not produce what you want, identify the specific gap and add instruction for that gap only.

The Debugging Mindset

When AI output is not what you wanted, treat it as a diagnostic problem rather than a failure. The question to ask is always the same: what information was missing from my prompt that would have changed this output? Most disappointing results map onto a small number of symptoms, and each symptom points at a specific omission. The table below is the fastest route from a bad output to the change that fixes it. Read it as a lookup rather than as a checklist to work through in order.

The gap between what you meant and what the model understood is the whole game. Close that gap and the output improves. That is why these ten patterns share a single antidote: assume nothing about what the model knows, and be relentlessly specific about what you want. Vagueness, overloading, missing context, weak format specification, absent guardrails, unrealistic expectations, inconsistency, misread capabilities, unexplained jargon, and unnecessary complexity account for the great majority of prompting failures, and specificity answers all ten.

Florian now gets useful competitive analysis outputs most of the time. His prompts are longer but more precise. He provides context, specifies format, and iterates. The tool did not change. His approach did.

Anti-Patterns

The patterns above describe how prompts go wrong. These describe how the diagnosis itself goes wrong, which is the failure mode that survives once you know the list.

• Using the debugging table as a lookup rather than a diagnosis. The symptom you notice and the omission that caused it are not always the same thing. Output that reads as too formal is sometimes a tone problem and sometimes a missing audience, and applying the tone fix to an audience problem produces a friendlier version of the wrong document. Ask what information was missing before you reach for the remedy.
• Changing several things at once. If you add context, tighten the format and shift the tone in a single revision, you learn nothing about which change mattered. The habit that builds skill is the one the lesson recommends for prompts generally: identify the specific gap and add instruction for that gap only.
• Answering weak output with more words. The instinct when a prompt underperforms is to make it longer, but length is not the variable. Elaborate instructions with nested conditions can confuse a model exactly as they would confuse a person, and many long prompts are verbose restatements of a simple one. Clarity is the fix; volume is a substitute for it.
• Blaming capability for a prompting problem, or a prompting problem for capability. Both errors cost you. Concluding that the tool cannot do this when the prompt was underspecified means abandoning a task that would have worked. Iterating for an hour against a knowledge cutoff, a calculation, or a request for genuine novelty means refining a request that no wording will rescue.
• Keeping the fix in your head instead of in the prompt. If you make the same manual edit to every output, that edit is a missing instruction, not a chore. Where the correction concerns style or structure and has to hold across repeated use or across a team, an example carries it more reliably than a description.

Practice Prompts

Each of these works on prompts you have already written, because the point is to see your own patterns rather than to study someone else's.

• Diagnose a disappointment. Take the last AI output that let you down. Write one sentence naming what information was missing from your prompt that would have changed it. Add only that, run it again, and see whether your diagnosis was right.
• Specify a routine request. Pick a prompt you use often and rewrite it with an audience, a purpose, a format, a length and a focus. Run the old and new versions side by side and note which parts of the improvement came from which addition.
• Hunt your own vague adjectives. Search your recent prompts for "good," "interesting," "relevant," "thorough," and "engaging." Replace each with the behavior you actually want, in the way that "use conversational tone, include one surprising statistic, and provide actionable tips" replaces "make it engaging."
• Break up an overloaded request. Find a prompt that asks for several deliverables at once. List the tasks it contains, decide which three matter most, and run only the first as its own prompt. Compare the result with what the combined version produced.
• Write a reusable context block. For a conversation you start repeatedly, draft the who, what and why an outsider would need, then apply the context audit: could someone who knows nothing about your situation act on this? Keep it where you can paste it.
• Show instead of tell. Ask for something in your own voice twice, once by describing the style and once by pasting an example and asking for the same tone. Decide which version you would have to edit less.
• Strip the jargon. Take a prompt containing your internal acronyms and rewrite it with the actual figures and situations spelled out, the way naming an acquisition cost and a lifetime value beats naming the ratio.

Reflection

Think about the last time you decided that AI was not useful for a particular task. Reconstruct the prompt you used. Would a capable colleague, given only those words and no knowledge of your situation, have produced what you were hoping for? If not, the conclusion you drew was about your prompt rather than about the tool, and it may have quietly ruled out work you could be doing. Now ask the harder version of the question: which of the ten patterns is your personal default? Most people have one they return to under time pressure, usually vagueness or overloading, because both feel efficient in the moment. Finally, consider what you currently fix by hand after every generation. That repeated correction is information about an instruction you have never written down.

Glossary

• Anti-pattern: A mistake that feels reasonable while you are making it and reliably fails in practice, which is what makes it worth learning to recognize rather than simply being told about.
• Context audit: The check applied before sending a prompt, asking whether someone who knows nothing about your situation could act on it, and adding the facts an outsider would lack.
• Few-shot prompting: Anchoring output to a concrete reference by providing one or more examples of what you want, which holds style and format more reliably than describing them.
• Guardrails: Explicit constraints on length, depth, structure, and what to include or exclude, which keep output inside the territory you care about.
• Format specification: Stating the shape the output must take, such as a table, a set number of bullet points, or a single paragraph, rather than accepting the model's default style.
• Overloaded prompt: A request carrying more tasks or constraints than can be satisfied together, so the model sets its own priorities and returns a compromise instead of a result.
• Knowledge cutoff: The limit beyond which a model has no information about events, which is why questions about recent developments cannot be fixed by better wording.
• Hallucination: The confident assertion of false information, a structural property of systems that synthesize from patterns rather than retrieve verified facts.

Related Lessons

These patterns sit alongside several other lessons in the sequence. Anatomy of an Effective Prompt and Core Prompting Patterns cover the constructive side, showing what a well-built prompt contains rather than what a broken one lacks, and they are the natural place to go if you find yourself rebuilding several prompts at once. Iterative Refinement develops the follow-up technique this lesson only sketches, which matters most if your diagnosis is sound but your second attempts are not converging. Critical Evaluation Framework takes over where debugging stops, addressing how to judge whether an output you now consider good is actually correct.

Closing

What changed for Florian was not his tooling and not his expectations. It was that he stopped reading a disappointing output as a verdict on the technology and started reading it as evidence about his own instructions. That shift is available immediately and costs nothing. The ten patterns here are worth knowing individually, but they collapse into a single working habit: assume nothing about what the model knows, be specific about what you want, and when the result disappoints you, ask what you left out before you ask what is wrong with the tool. If you take one practice into your week, make it the diagnostic question. It converts a frustrating session into a correction you can apply the next time.

Key Takeaways

• Vague prompts produce vague outputs. Specify audience, purpose, format, length, and focus, and strip out words like "good" or "engaging" that leave the judgment to the model. If a colleague would need to ask clarifying questions, so does the AI.
• Break complex tasks into sequential prompts. Overloaded prompts produce compromises. Pick the three things that matter most, and turn the rest into follow-ups you can verify one at a time.
• Prime each conversation with context. The AI starts fresh every time. Give it the who, what, and why, and describe your situation rather than labeling it with internal jargon.
• Specify format and set guardrails. State length, structure, depth, and what to exclude. Explicit boundaries prevent the wandering that makes otherwise correct output unusable.
• Show rather than tell for style and consistency. A concrete example, or a template every output must follow, anchors results far more reliably than a description of the style.
• Iteration is normal, not failure. First drafts need refinement, and the strongest results typically come after two or three targeted passes. Budget the time rather than treating it as a sign the tool failed.
• Diagnose before concluding the tool does not work. Most "AI doesn't work for this" conclusions are really "my prompt didn't have what the task needed," and a few are genuine capability limits worth recognizing as such.

Frequently Asked Questions

What is the most common prompt engineering mistake? Being too vague. People assume the model knows what they want without stating it, so instead of "Summarize this," the prompt should say something closer to Extract the three key findings from this research paper and explain why each one is significant for the healthcare industry. Specificity fixes most poor results, which is why it is worth exhausting before you conclude anything about the tool.

How can I tell if my prompt has a context problem? If the output is vague, generic, or built on default assumptions that do not match your situation, the prompt probably lacks context. Test it the way you would test a briefing: would a smart colleague need to ask clarifying questions before starting? If they would, so does the model, and the questions they would ask are exactly the facts to add.

What should I do if I think the model is hallucinating? Verify the facts independently first rather than trusting what it says. If the information is provably wrong, supply the correct information in your next prompt and ask it to redo the task. You can also ask it to cite sources or explain its reasoning, which sometimes reveals where it is uncertain. Treat this as routine for anything you intend to rely on.

Are longer prompts better? Not by themselves. Adding specificity helps; adding words does not. Past a point, extra constraints reduce quality, because the model has to guess which requirements matter and ends up doing everything superficially. Start simple, and when the simple version falls short, identify the specific gap and add instruction for that gap alone.

How many rounds of iteration should I expect? Enough that you should plan for them rather than treat them as overrun. The strongest results usually arrive after two or three passes in which you read the output, form a view about what is wrong, and refine. If several passes have not moved the result, the problem is more likely in the prompt's framing than in the wording of your feedback.

Does any of this change if my colleague uses a different tool? The specific defaults differ, so tone, length and format habits vary between tools. The patterns do not. Vagueness, overloading, missing context, absent guardrails and unexplained jargon degrade output everywhere, because they are all versions of the same problem: the words you sent did not contain what the task required.