Interactive eBook BuilderInteractive eBook Builder
$59 · unlimited access, unlimited eBooksGet it for $59

· By · 10 min read

Negative Prompting: What to Tell AI Not to Do

promptsaiwritingebook

Negative prompting means telling the model what to leave out—invented facts, the wrong audience, a format you do not want, or content that does not belong in this chapter—so the result stays inside your boundaries. Official guidance supports the idea, with an important nuance. OpenAI’s prompt engineering guide says the instructions section should cover what the model should do and what it should never do. Google’s prompt design strategies say you can specify constraints on how the model reads the prompt or generates a response, including what to do and not to do. Anthropic’s prompting best practices advise, for output format in particular, that it is often more effective to tell Claude what to do instead of what not to do. The article’s own example is to prefer “Your response should be composed of smoothly flowing prose paragraphs” over “Do not use markdown in your response.”

Those pieces of advice are compatible if you use each one for the right job. When you want a shape, describe the shape. When you want to forbid a specific failure—invented sources, a medical claim, a preamble, an item that was not in the notes—say so clearly.

For eBook and workbook projects, Interactive eBook Builder is a $59 one-time Notion AI template with Spec Builder, Drafter, Polisher and SEO BOT already set up, so limits can live in the standing instructions instead of being appended in a panic after a bad draft. Notion AI access and usage are separate from the purchase. The examples below are teaching prompts written for this article, not paid Builder prompts and not recorded model outputs.

What counts as a useful negative instruction?

A useful negative instruction names a behavior you can observe in the output. “Do not invent a statistic that is not in the notes” can be checked. “Do not be boring” cannot. “Do not add a preamble such as 'Here is your chapter'” can be checked. “Don’t sound like AI” usually cannot, unless you define the tells you care about, such as a list of banned stock phrases.

Good candidates for eBook and workbook work include:

Anthropic also recommends explaining why a constraint matters, because the model can generalize from the reason. “Never use ellipses” is weaker, in their example, than explaining that a text-to-speech engine will have to read the response. For a book, “Do not add supplement advice; this chapter is only about a 45-minute weekday routine” is stronger than a bare “no biohacking.”

When a positive instruction should replace the ban

If you are really arguing about format, a ban is often a clue that you have not described the output. Anthropic’s markdown example is the pattern. Instead of “Do not write an essay,” say “Return six checkbox items, then one reminder sentence.” Instead of “Don’t ramble,” say “250 to 350 words, then one example.” Instead of “No bullet-point soup in the teaching section,” say “Write this section in short prose paragraphs, and reserve lists for the example steps at the end.”

Google likewise treats response format as something you can specify directly: a table, a list, a paragraph, a set of keywords. If the model has a clear target, you need fewer bans.

A useful test is to try deleting the negative line. If the remaining brief still describes a result you can inspect, the ban may be optional. If deleting it would let the exact miss you care about back in, keep it.

A before-and-after for chapter notes

A vague prompt with a vague ban:

Write a chapter section on morning routines for freelancers. Don't be
generic and don't add anything weird.

That ban cannot be checked. A tighter version uses a positive format and a short, specific ban list:

Draft a 250 to 350-word section, "Choose three actions, not a full
lifestyle change," for new freelancers who have 45 minutes before the
first client call, three days a week.

Use only these notes:
- The candidate actions are: review today's calendar, write the first
  task in one sentence, and prepare the workspace for the first call.
- The reader may swap one action, but the trio must stay under 45 minutes.

Write short prose paragraphs, then end with the three actions as numbered
steps and a time estimate for each.

Do not:
- add supplements, 5 a.m. schedules, or new apps
- invent research or success rates
- open with a founder story
- add a preamble such as "Here is the section you requested"

The “do not” list is short, concrete, and based on likely misses. The format lines do most of the shaping. If the draft still adds a meditation timer that was not in the notes, you can point to a specific rule that failed and tighten that one line. You cannot do that with “don’t add anything weird.”

A before-and-after for a workbook page

Workbook pages often fail by growing extra activity. The model adds scoring, personality labels, or a second exercise you will not have space for.

A weak version:

Make a worksheet for the revision-limit chapter. Don't make it too
complicated.

A stronger version:

Create one five-minute workbook page for the chapter on stating a
revision limit in the kickoff email.

Format:
1. A one-sentence instruction
2. A line for the reader's revision limit
3. A line for the one-sentence email they will send
4. One backup question: "If a client pushes back, what will you repeat?"

Do not add scoring, personality labels, extra exercises, or legal wording.
Use only the chapter notes, and leave writing space on each line.

Here the positive format is doing the heavy lifting. The negative list is a fence around failures you can predict. If you later see a new miss, such as a motivational quote at the bottom, add that one item after you have seen it, rather than trying to ban every possible decoration in advance.

How to keep a ban list from taking over the prompt

A standing “never” list can grow until nobody reads it, including you. Keep it in the system or template section, and keep it short. Add a new item only when a miss has appeared more than once, or when a claim would be harmful even once, such as a medical or legal invention.

Put the bans next to the related positive rule when you can: “Stay inside the notes. Do not invent rates, studies or quotes.” That pairing is closer to OpenAI’s “what to do and what never to do” than a second page of prohibitions at the end of the prompt.

If you work in Claude, and the problem is formatting, try Anthropic’s positive-first approach before you add another “do not.” If you work in Gemini, Google’s constraints section officially includes both do and do not; still describe the output you want. If you work in ChatGPT or the OpenAI API, put the never-do rules in the instructions or standing brief, not only in a passing remark at the bottom of a long request.

For related techniques, delimiters help the model see which block is notes and which block is the request, which prevents a different kind of overreach: treating your instructions as source material, or your notes as optional. And if a ban is not working, our guide to improving a prompt when the answer is generic is the revision loop to use next.

Where this fits a longer workflow

In Interactive eBook Builder, limits can sit in the standing instructions for Spec Builder, Drafter and Polisher so each stage inherits the same boundaries: no invented sources, no unsupported income claims, no format the book does not use. That is more reliable than remembering to paste a ban list into every chat. SEO BOT can later publish supporting articles to the Auto-Blogger blog, each linking back to your eBook, site and brand; those articles need the same source and claim limits, or the blog will undo the discipline of the book. You can keep the supplied setup or adapt it. You still review the work, especially wherever a “do not invent this” rule meets a fluent paragraph.

FAQ

Is negative prompting the same as negative prompts in image models?

The idea is related—you are stating what to avoid—but the habits are not identical. Image tools often have a dedicated negative-prompt field. In writing tools, the same job is usually a constraint inside the instructions. Do not assume that a long comma-separated ban list from an image workflow will transfer cleanly to a chapter brief.

Why does the model still do the thing I banned?

Sometimes the ban is vague, or it fights a stronger format instruction, or the notes themselves invited the miss. Sometimes the model simply fails the instruction; responses are not deterministic. Tighten the rule, add a short reason, show an example of the output without the banned behavior, and test again. If the miss is about format, rewrite the request as a positive shape.

Should I put all bans in capital letters?

No official guide in the sources above requires that, and shouting at the model is not a substitute for a clear rule. A short, specific line in the instructions is enough. If you have many equally “critical” caps, the emphasis stops meaning anything.

Can I ban AI tells such as “delve” or “in the ever-evolving landscape”?

You can ban a short list of phrases you actually see, and you can describe the voice you want instead. A long list of every possible tell is hard to maintain and still misses new ones. Voice samples, as in few-shot prompting, are often more effective than a dictionary of prohibitions.

Before you hit send

Ship niche eBooks with agents already built.

Get the Interactive eBook Builder for $59