AGENCYSCRIPT
CoursesEnterpriseBlog
πŸ‘‘FoundersSign inJoin Waitlist
AGENCYSCRIPT

Governed Certification Framework

The operating system for AI-enabled agency building. Certify judgment under constraint. Standards over scale. Governance over shortcuts.

Stay informed

Governance updates, certification insights, and industry standards.

Products

  • Platform
  • Certification
  • Launch Program
  • Vault
  • The Book

Certification

  • Foundation (AS-F)
  • Operator (AS-O)
  • Architect (AS-A)
  • Principal (AS-P)

Resources

  • Blog
  • Verify Credential
  • Enterprise
  • Partners
  • Pricing

Company

  • About
  • Contact
  • Careers
  • Press
Β© 2026 Agency Script, Inc.Β·
Privacy PolicyTerms of ServiceCertification AgreementSecurity

Standards over scale. Judgment over volume. Governance over shortcuts.

On This Page

What Negative Prompting Actually IsTwo mechanisms, one ideaConstraints versus exclusionsWhy Exclusions Are Harder Than They LookSalience can backfireNegatives compete with positivesThe fix is often to invertWhen to Reach for a Negative PromptStrong cases for exclusionWeak cases for exclusionHow to Write Constraints a Model Can FollowBe specific and observableGroup and label your prohibitionsPair the prohibition with the reason or the alternativeTest the inversion firstNegative Prompting in Image GenerationCommon negative keywordsCalibrate strengthKeep it terseMeasuring Whether Your Constraints WorkRun an A/B comparisonWatch for overcorrectionKeep a tested constraint libraryFrequently Asked QuestionsDoes telling a model not to do something make it more likely to do it?Is negative prompting the same in image and text models?How many negative constraints can I include before quality drops?When should I use a positive instruction instead?Key Takeaways
Home/Blog/What to Tell a Model It Should Never Do
General

What to Tell a Model It Should Never Do

A

Agency Script Editorial

Editorial Team

Β·January 10, 2023Β·8 min read
negative promptingnegative prompting guidenegative prompting guideprompt engineering

Most prompt advice tells you what to ask for. Far less of it tells you what to forbid. Yet much of the gap between a mediocre output and a usable one comes down to constraints: the formats you refuse to accept, the tones you rule out, the topics you keep off the table. Telling a model what not to do is a skill in its own right, and it behaves differently from the positive instructions that get all the attention.

Negative prompting is the practice of shaping output by specifying what to exclude rather than only what to include. It shows up in two distinct worlds. In image generation, a dedicated negative prompt field steers a diffusion model away from artifacts like extra fingers or watermarks. In text and chat models, negative instructions live inside the ordinary prompt as constraints: do not invent statistics, never address the reader by name, avoid hedging language. Both rely on the same intuition, but they work through different mechanisms and reward different habits.

This guide is the structured overview for someone who wants to use exclusions deliberately rather than by trial and error. We will cover where negative prompting genuinely helps, where it quietly backfires, and how to phrase constraints so a model can actually follow them. Treat it as the map; the sibling articles linked throughout go deeper on specific terrain.

What Negative Prompting Actually Is

A negative prompt is any instruction whose purpose is to suppress, exclude, or prevent something in the output. That something might be a visual element, a word, a tone, a format, or an entire class of behavior.

Two mechanisms, one idea

In diffusion image models, the negative prompt is a separate conditioning signal. The model is steered toward your positive prompt and away from the negative one during sampling. The exclusion is mathematical, not conversational, which is why image negative prompts tend to be terse keyword lists rather than sentences.

In language models, there is no separate channel. A negative instruction is read alongside everything else, and the model has to interpret it. That difference matters: a text model can technically reason about your prohibition, but it can also misread, overgeneralize, or fixate on the very thing you told it to avoid.

Constraints versus exclusions

It helps to separate two things people lump together. A constraint bounds the output ("keep it under 100 words"). An exclusion removes a possibility ("do not mention pricing"). Both are negative in spirit, and the rest of this guide treats them as a family. For the absolute basics framed for a first-time reader, start with Negative Prompting: A Beginner's Guide.

Why Exclusions Are Harder Than They Look

The classic warning is that telling someone not to think of a pink elephant guarantees they think of one. Language models are not immune to this. A poorly placed negative can prime the exact content you wanted gone.

Salience can backfire

When you write "do not use the word synergy," you have now placed the word synergy in the context window. For some models and some phrasings, this raises rather than lowers the chance it appears. The effect is inconsistent across models and versions, but it is real enough that you should test rather than assume.

Negatives compete with positives

Every prohibition adds cognitive load. A prompt with twelve things to avoid and three things to do often produces output that satisfies the avoidances while neglecting the goals. Models have finite attention, and a wall of "do not" crowds out the affirmative direction the output actually needs.

The fix is often to invert

The most reliable repair is to convert a negative into a positive whenever you can. Instead of "do not be formal," say "write casually, like a text to a friend." Instead of "avoid long sentences," say "keep most sentences under fifteen words." Inversion gives the model a target to move toward rather than a void to avoid.

When to Reach for a Negative Prompt

Not every problem is a negative-prompting problem. Knowing when exclusions are the right tool saves you from over-constraining.

Strong cases for exclusion

  • Removing recurring artifacts you have seen across multiple outputs (a stock phrase, a closing line, a visual glitch)
  • Enforcing compliance boundaries, such as never giving medical or legal advice
  • Suppressing categories that have no clean positive equivalent, like "do not fabricate sources"
  • Image generation, where the negative field is purpose-built for exclusion

Weak cases for exclusion

  • Anything you can state positively with equal precision
  • Vague aesthetic preferences that the model cannot operationalize ("don't be boring")
  • Long lists of edge cases that would be better handled by an example

For concrete before-and-after scenarios that show this judgment in action, see Negative Prompting: Real-World Examples and Use Cases.

How to Write Constraints a Model Can Follow

Phrasing is where most negative prompts succeed or fail. A few patterns separate constraints that hold from ones that get ignored.

Be specific and observable

"Do not be verbose" is unenforceable because verbosity is subjective. "Do not exceed three sentences per paragraph" is checkable. Whenever possible, write exclusions a person could grade with a ruler.

Group and label your prohibitions

Burying "never cite a statistic without a source" in the middle of a paragraph makes it easy to miss. Put hard constraints under a clear heading like "Rules" or "Never do the following," and keep the list short. Structure raises adherence.

Pair the prohibition with the reason or the alternative

Models follow constraints better when they understand the intent. "Do not use bullet points; this is a narrative essay" outperforms "do not use bullet points" alone, because the alternative is now obvious. The step-by-step build of this technique lives in A Step-by-Step Approach to Negative Prompting.

Test the inversion first

Before adding a negative, ask whether a positive instruction would do the job. Reserve true negatives for cases where no positive phrasing captures what you mean.

Negative Prompting in Image Generation

Because the image case has its own dedicated field, it deserves separate treatment. Here the negative prompt is less about behavior and more about visual cleanup.

Common negative keywords

Practitioners maintain reusable negative lists: blurry, low quality, deformed, extra limbs, watermark, text, jpeg artifacts. These get appended to most prompts as a baseline hygiene layer.

Calibrate strength

Many interfaces let you weight the negative prompt. Too weak and artifacts slip through; too strong and you start losing legitimate detail. The right weight is found empirically, and it differs by model.

Keep it terse

Unlike text models, diffusion negatives reward keyword density over grammar. Full sentences waste tokens and rarely help.

Measuring Whether Your Constraints Work

A negative prompt is only useful if it changes outputs in the direction you intended without collateral damage.

Run an A/B comparison

Generate several outputs with and without the constraint. If the prohibited element disappears and quality holds, keep it. If quality drops or the model overcorrects, revise. The Case Study: Negative Prompting in Practice walks through exactly this kind of measured iteration.

Watch for overcorrection

A constraint against jargon can produce prose so plain it reads as condescending. Always check that the cure is not worse than the disease.

Keep a tested constraint library

Once a negative reliably works for your use case, save it. Reusable, vetted constraints are more valuable than clever one-offs.

Frequently Asked Questions

Does telling a model not to do something make it more likely to do it?

Sometimes, especially with vague or heavily repeated prohibitions, because naming a concept places it in context. The risk is real but not universal. The defense is to prefer positive phrasing, keep negatives short and specific, and test outputs rather than trusting that the instruction was obeyed.

Is negative prompting the same in image and text models?

No. Image models often have a dedicated negative prompt field that steers sampling away from your keywords through a separate conditioning signal. Text models read negatives as ordinary instructions they must interpret. The intent is similar, but image negatives reward terse keyword lists while text negatives reward structured, reasoned phrasing.

How many negative constraints can I include before quality drops?

There is no fixed number, but adherence degrades as the list grows because constraints compete with your positive goals for the model's attention. A practical ceiling for most text prompts is a short, grouped list of the few prohibitions that matter most. If you find yourself listing many, consider whether an example or a positive instruction would replace several of them.

When should I use a positive instruction instead?

Almost always, if a positive version expresses the same idea with equal precision. Reserve true negatives for compliance boundaries, fabrication prevention, and cases where no clean positive equivalent exists. Inverting "do not X" into "do Y" gives the model a target rather than a void.

Key Takeaways

  • Negative prompting shapes output by specifying what to exclude, and it operates differently in image models (a dedicated field) than in text models (interpreted instructions).
  • Exclusions are harder than they look: vague or repeated prohibitions can prime the unwanted content and crowd out your actual goals.
  • Convert negatives to positives whenever a positive phrasing carries the same meaning with equal precision.
  • Write constraints that are specific, observable, grouped under a clear heading, and paired with a reason or alternative.
  • Always test with an A/B comparison and watch for overcorrection before locking a constraint into your reusable library.

Search Articles

Categories

OperationsSalesDeliveryGovernance

Popular Tags

prompt engineeringai fundamentalsai toolsthe difference between AIMLagency operationsagency growthenterprise sales

Share Article

A

Agency Script Editorial

Editorial Team

The Agency Script editorial team delivers operational insights on AI delivery, certification, and governance for modern agency operators.

Related Articles

General

Rolling Out AI Hallucinations Across a Team

Most teams discover AI hallucinations the hard way β€” a confident-sounding wrong answer makes it into a client deliverable, a legal brief, or a published report. The damage isn't just to the output; it

A
Agency Script Editorial
June 1, 2026Β·11 min read
General

Case Study: Large Language Models in Practice

Most teams that fail with large language models don't fail because the technology doesn't work. They fail because they treat deployment as a one-time event rather than a discipline β€” pick a model, wri

A
Agency Script Editorial
June 1, 2026Β·11 min read
General

Thirty-Second Wins Breed False Confidence With LLMs

Working with large language models is deceptively easy to start and surprisingly hard to do well. You can get a useful output in thirty seconds, which creates a false confidence that compounds over ti

A
Agency Script Editorial
June 1, 2026Β·10 min read

Ready to certify your AI capability?

Join the professionals building governed, repeatable AI delivery systems.

Explore Certification