Saying "What You Want" Isn't Enough: Include "Don't Do This" in Your AI Prompts

Last week, I asked an AI to write a product copy draft. The first version was complete, but the screen was filled with clichés like "ultimate experience," "empowering users," and "creating a closed loop." I deleted them, it learned, and I repeated this cycle three times. On the fourth attempt, I added one line to the prompt:

> Do not use: empower, closed loop, ultimate, seamless, create.

This time, the output was ready to use immediately.

Many prompts fail not because the AI isn't smart enough, but because you only told it "what to do" without telling it "what to avoid."

How to Incorporate "Don't Do This"

Here are three approaches, ranging from light to heavy constraints.

**1. Direct Ban.** Explicitly state in the task: "Do not use word X" or "Do not use structure X." This works well for habitual errors. For Chinese copywriting, ban phrases like "It is worth noting that" or "In conclusion"; for coding, ban "starting variables with `var`."

**2. Provide Counter-Examples.** Paste a bad sample directly:

> Here is the style we do NOT want: ...

> Reason: Too salesy, full of adjectives, lacks specific numbers.

One concrete counter-example is far more effective than ten vague instructions to "be more professional."

**3. Create a "Don'ts" List.** Compile recurring prohibitions into a small file. At the start of the task, include a reference like: "Read `./donts.txt` first; do not use any of these styles." This is ideal for teams, long-term tasks spanning multiple sessions, or when handing off rules to other models or colleagues.

A More Concrete Example

I asked an AI to revise a server alert email template. Three attempts failed: the reminders were too wordy, and it included unnecessary pleasantries at the beginning. On the third try, I changed my approach:

> Refer to the previous version that needs fixing (anonymized below): The issues are the polite opening and lack of a clear call to action.

> New requirements: State what happened in the first sentence, provide the command in the second, and keep it under five lines.

> Do not use: Hello, Friendly Reminder, Thank you for your support.

The next version was correct. By contrasting positive and negative examples, the AI understood much better than it would have from ten sentences of adjectives.

When to Use This

- The same type of error has occurred more than twice. Ban it on the second occurrence; don't wait for a third.

- The deliverable has clear red lines: brand terms, prohibited words, or compliance requirements.

- The task is being handed off to a new session, colleague, or model. Old habits are most likely to resurface during handovers.

When NOT to Use This

- You don't yet know what you *don't* want. If requirements are still being explored, let the AI run freely first. Once it produces a draft, identify the flaws and then apply bans.

- During brainstorming, naming, or divergent creative thinking. Too many negative constraints make the AI conservative, resulting in safe but mediocre answers. During the divergent phase, zero to one hard constraint is enough; save the rest for the convergent phase.

A 30-Second Checklist Before You Start

1. Where did it fail last time? That specific issue becomes your first "Don't."

2. Are there more than five prohibited words? If so, create a separate list file instead of cluttering the main prompt.

3. Is this task divergent or convergent? If divergent, write fewer "Don'ts."

Common Pitfalls

- **More negative constraints do not equal more stability.** In image generation, if the negative prompt exceeds seven or eight items, image quality often deteriorates; the same applies to text. When there are too many bans, the model tends to prioritize the most recent one, rendering earlier ones ineffective. This can even lead to an ironic effect—the more you say "don't use 'ultimate'," the more "ultimate" appears in the draft.

- **Banning without demonstrating causes the AI to spin its wheels.** The correct approach is "Don't do X, do Y instead": for every prohibition, provide an alternative. Simply stating what not to do leaves the AI unsure of what you actually want to see.

- **Anonymize counter-examples.** Before using real deployment data or customer information as negative examples, strip out names and numbers. The cost of a counter-example leaking is far higher than the benefit of a good example.

Including "Don't Do This" in your prompts is essentially communicating your baseline standards to the executor in advance. These three words are often more effective than ten sentences of praise.