Quick answer
Negative constraints in prompting are explicit rules telling an AI what not to produce — banned words, formats, or behaviours. On their own they are frequently ignored. Pair every ban with the positive target you want instead, cap the list at roughly five per prompt, and store permanent rules in your account settings.
Negative constraints in prompting are the rules that tell an AI what not to produce: banned words, banned formats, banned behaviours. Most professionals never write one. They generate a draft, wince at an opening line about a rapidly evolving landscape, and then spend fifteen minutes deleting by hand what a single sentence could have prevented.
Here is the part most guides skip. A ban on its own often makes things worse. Type “do not use bullet points” and you will frequently get bullet points anyway. That is not the model being difficult with you — it is a predictable failure with a fix that takes about four extra words. Below is the working method, in the order you will need it:
- Why a bare “do not” gets ignored, and what both OpenAI and Anthropic recommend instead
- The Fence-and-Path Method — how to write a ban that actually holds
- The four kinds of rules worth setting at work
- Five copy-paste templates for the documents you already produce
- Where to store your rules so you never type them twice
None of this needs coding, an API, or a paid plan. If the wider vocabulary here is new to you, our guide to what prompt engineering actually means is a useful companion piece, though you can follow everything below without it.
Will this work for you?
What Negative Constraints in Prompting Really Are
A negative constraint is an explicit instruction stating what the AI must not include, in its content, its wording, its structure, or its behaviour. Positive instructions describe the deliverable you want. Negative constraints describe the failures you refuse to accept. Used together, they narrow the output before generation rather than after it.
In practice, professionals reach for them to stop four recurring irritations:
- Filler openings — the throat-clearing sentence before any real content arrives
- Buzzwords — “tapestry”, “transformative”, “robust”, the vocabulary that reads as machine-written
- Wrong shape — bullets when you asked for prose, an essay when you needed a table
- Invented detail — confident claims that were nowhere in the file you uploaded
Why chat AI differs from image tools
Search for “negative prompting” and most of what surfaces is about image generators, where a separate box exists purely for exclusions like “blurry, extra fingers, watermark”. That box does not exist in ChatGPT, Claude, Gemini, or Copilot. Your exclusions live in the same message as everything else, which changes how they behave.
| Where you write it | Image generators | Chat AI tools |
|---|---|---|
| Location | A dedicated negative prompt field | The same message as your instructions |
| How it is read | A separate exclusion list | One instruction among many |
| What it needs | Nothing else | A stated target alongside it |
That last row is the whole article in miniature, and it is why copying image-generator advice into a chat window disappoints people. Exclusions in chat need company. Getting the wanted half right first is covered in our guide to specifying format so AI output is immediately usable.
Why “Do Not” Rules Get Ignored
Negative constraints fail for one structural reason: a bare prohibition tells the model where it cannot go without telling it where to arrive. It still has to produce something, so it fills the gap with its default behaviour — which is usually the exact habit you were trying to stop.
✕ Ban with no target
"Summarise this report. Do not use bullet points. Do not delve into background. Do not use corporate jargon."
✓ Target, then ban
"Summarise this report in three short paragraphs, opening with the cost variance. Do not use bullet points."
The second version is shorter and works far more reliably, because “three short paragraphs, opening with the cost variance” gives the model somewhere to go. The ban is now a boundary on a defined route rather than a fence in an empty field.
A ban with no destination leaves the model to fall back on its defaults.
This is not a workaround someone discovered on a forum. Both major AI companies document it as standard practice. OpenAI’s best practices for prompt engineering lists it as a numbered rule: instead of only saying what not to do, say what to do instead. Anthropic’s prompting best practices for Claude opens its formatting guidance with the same instruction, contrasting a flat refusal of markdown against a positive request for flowing prose paragraphs.
When the two largest labs independently publish the same correction, it stops being a trick and starts being the documented way the tools are meant to be used.
The Fence-and-Path Method
The Fence-and-Path Method is the approach we teach at PromptPeakAI for writing negative constraints in prompting that actually hold: every fence you build must have a path running through it. The fence is your ban. The path is the destination you want instead. A fence without a path is a request for the model to guess, and it guesses badly.
Every rule needs both halves — the destination and the boundary around it.
Building one takes four passes, and the first is the one people skip.
Name what you keep deleting
Open the last three documents you fixed by hand and look at your own edits. Whatever you removed twice is your real ban list. This beats copying someone else’s list of forbidden words, because your irritations are specific to your job and your reader.
Write the path before the fence
For each ban, finish this sentence: “instead, do this.” No answer means you have a preference, not a rule, and it belongs in the bin rather than the prompt. Writing the destination first also stops you banning something you actually need.
Cap the list at five
Beyond roughly five bans in one message, quality tends to slide the other way: answers get shorter, hedged, and oddly cautious as the model tries to satisfy every restriction at once. Keep the five that cost you the most editing time. The rest belong in your account settings, which we get to shortly.
Test on one real document
Run the rule against a task you have already done manually, so you have something honest to compare against. If a banned word survives, the fix is almost never a sterner ban — it is a clearer path.
Task: [what you want produced, in one line] Path — do this: 1) Structure: [exact shape, e.g. three paragraphs then a two-column table] 2) Voice: [e.g. plain business English, concrete nouns, active verbs] 3) Anchor: [what the output must open with or centre on] Fence — do NOT do this: 1) Do NOT open with a scene-setting or context sentence. 2) Do NOT use the words: delve, tapestry, robust, transformative, leverage. 3) Do NOT use passive voice. 4) Do NOT include anything not present in the source material. Source material: [paste your document, notes, or data here]
You should now have a prompt where every "do NOT" line has a matching "do this" line above it.
If your bans are mostly about voice rather than structure, showing beats telling — paste two paragraphs of your own writing as a sample instead. That technique is covered in our guide to few-shot prompting and matching your voice, and it pairs well with a short fence.
The Four Kinds of Bans
Rules that work at work fall into four groups. Naming the group first makes it obvious what the matching path should be, and stops you writing five versions of the same style rule while your structural problem goes unaddressed.
Four categories of workplace ban, each with the kind of path it needs.
| Category | The ban | The matching path |
|---|---|---|
| Content | No speculation about causes | Quote the exact figure from the file |
| Style | No buzzword vocabulary | Concrete nouns and active verbs |
| Structure | No bulleted lists | Three paragraphs then one table |
| Behaviour | No rewriting my sentences | Flag the issue in a comment |
The trap in style bans
Style is where most people start and where most people write their rules too vaguely. “Do not sound corporate” gives the model nothing to measure itself against. A named list of forbidden words does, because a word is either present or it is not.
✕ Unmeasurable
"Do not sound corporate or use too much jargon. Try to avoid buzzwords where possible."
✓ Measurable
"Write in plain business English. Do NOT use these words: tapestry, robust, transformative, holistic, elevate."
Note the other change there. “Try to avoid” and “where possible” read as preferences, and get treated as such. “Do NOT” reads as a boundary. Softening your own rule is the quietest way to have it ignored.
The one people forget
Behavioural bans define the model’s role rather than its output, and almost nobody writes them. “Do NOT rewrite my sentences — flag anything unclear and let me decide” turns an editor that quietly overwrites your voice into one that hands the decision back. It is the single most useful rule for anyone who cares how their own writing sounds. This is the Role element of the Logic System: Role, Goal, Context, Format, expressed as a boundary rather than a job title.
Five Templates for Common Documents
These are complete fence-and-path blocks for the tasks that generate the most rewriting — negative constraints in prompting, already paired with their targets. Replace the bracketed parts and paste. Each one caps the fence at four rules deliberately.
Executive summaries
The classic failure here is a summary that takes two sentences to start summarising. Pair it with our walkthrough on writing an executive summary of any document for the structure itself.
Task: Summarise the attached report for our leadership team. Path: Open with the single most important number and what changed. Then cover cost variance, milestone status, and blockers. Maximum 200 words. Plain business English. Fence: 1) Do NOT write a scene-setting opening sentence. 2) Do NOT use the words: robust, transformative, holistic, tapestry. 3) Do NOT use passive voice. 4) Do NOT include any figure that is not in the attached report.
Client and sales emails
Outreach is where AI writing is most obvious to the recipient, usually because of manufactured warmth. Our guide to writing cold outreach emails that get replies covers the persuasion side.
Task: Draft a follow-up email to [client] after [meeting or demo]. Path: Reference [the specific thing we discussed], answer [their open question], and propose two concrete times next week. Under 120 words. Sound like a colleague, not a brochure. Fence: 1) Do NOT open with "I hope this email finds you well" or similar. 2) Do NOT use exclamation marks. 3) Do NOT offer a discount, extension, or incentive I have not authorised. 4) Do NOT restate what they already know about their own business.
Procedures and SOPs
Procedure drafts fail by drifting into policy theory. The fence forces every line to be something a person can actually perform on a Tuesday morning.
Task: Write a standard operating procedure for [process]. Path: Numbered steps, one action per step, each starting with a verb. Every step names the role responsible and the deadline. Write for someone doing this on their first day. Fence: 1) Do NOT include background, rationale, or policy explanation. 2) Do NOT use vague qualifiers: as appropriate, properly, in a timely manner. 3) Do NOT combine two actions into one step. 4) Do NOT leave any step without a named owner.
Meeting recaps
Recaps bloat because the model treats every exchange as equally worth reporting. Our guide to turning a meeting transcript into actionable minutes goes deeper on the format.
Task: Turn the transcript below into a recap for people who missed the meeting. Path: Three sections — Decisions made, Actions with owner and due date, Open questions. Every action names one person. Fence: 1) Do NOT include small talk, scheduling chat, or side discussions. 2) Do NOT list an action without an owner and a date. 3) Do NOT summarise a topic that produced no decision or action. 4) Do NOT infer agreement that was not clearly stated. Transcript: [paste transcript here]
Data and analysis notes
Analysis is where an invented explanation does real damage, because a confident sentence about why revenue fell is very hard to spot as fiction.
Task: Analyse the attached sheet and write findings for [audience]. Path: Identify the three largest changes. For each, state the exact figures, the period compared, and what it means for [team or decision]. Fence: 1) Do NOT explain WHY a number changed unless the reason is stated in the file. 2) Do NOT round any figure. 3) Do NOT state general truths such as "revenue growth is important". 4) Do NOT include a finding you cannot point to a cell for.
If you would rather start from a ready-made set than build your own, the banned-word and constraint blocks in our free AI work templates cover the same five document types.
Where Your Rules Should Live
Retyping the same negative constraints into every chat is the fastest way to abandon the habit. Rules belong at the highest level that still makes sense for them, and every major tool now has three levels to choose from.
Move a rule up a level each time you find yourself typing it again.
Your banned-word list is almost always an account-level rule. It never changes between documents, so typing it repeatedly is pure waste. In ChatGPT this is Custom Instructions, and it is worth knowing that OpenAI caps that field at 1,500 characters on the free and Go plans, with more room on paid ones — enough for a tight list, not enough for a manifesto. Our guide to setting up ChatGPT Custom Instructions for professional work walks through the setup.
The ChatGPT Settings screen open at Personalization, showing Custom Instructions with a short banned-word list and a paired positive instruction typed into the response field.
Crop to the settings panel only. Hide the account name, email, avatar, and the chat history sidebar.
File: chatgpt-custom-instructions-banned-words.png · Alt: "ChatGPT custom instructions showing negative constraints and a banned word list"Claude’s equivalent is per-project instructions, which suits people whose rules differ by client — a legal client and a marketing client can have opposite banned-word lists without either bleeding into the other. Our guide to using Claude Projects to organise your AI work covers project knowledge and instructions together.
A Claude Project's custom instructions panel containing a fence-and-path block, with the positive target above and four numbered do NOT rules below.
Crop to the instructions panel. Hide the workspace name, user name, and any project titles that identify a real client.
File: claude-project-instructions-fence-and-path.png · Alt: "Claude Project instructions holding negative constraints for professional writing"What Negative Constraints in Prompting Cannot Do
Three limits of negative constraints are worth knowing before you trust the technique with anything that matters, and one of them is a genuine risk rather than an inconvenience.
A ban is a formatting rule, not a security control.
Writing "do NOT share this data" changes the wording of the response. It does not change what the tool received, what gets stored, or what your provider's retention policy says. The only reliable control is not pasting the material in the first place.
- Bans do not stop invented detail. "Do NOT make anything up" reduces obvious speculation, but it cannot make the model know what it does not know. Verification stays your job.
- Bans cannot enforce judgement. No rule tells the model which of your three findings actually matters to your director this quarter.
The security point deserves its own picture, because “I told it not to” is a sentence people genuinely use to justify a paste they should not have made.
Your own drafts, published material, anonymised examples.
Client names, internal figures, anything identifying a colleague.
Credentials, personal data, unreleased financials, signed contracts.
On invented detail specifically, a fence helps at the margins and no further. It will stop a model padding an analysis with plausible-sounding causes, and it will not stop a confident wrong citation. The mechanics of that are covered in our explanation of why AI hallucination happens and how to catch it. Most experienced analysts will tell you the same thing: a constraint changes what the output looks like, and only a human changes whether it is true.
When Your Rules Stop Working
Negative constraints degrade for four predictable reasons, and each has a fix that takes under a minute. Work down this table in order — the first row solves most cases.
| What you see | Why it happens | The fix |
|---|---|---|
| The banned word appears anyway | No path was written | Add the "instead, do this" line |
| Answers turn short and hedged | Too many rules at once | Cut the fence back to five |
| The rule fades mid-conversation | It scrolled out of context | Restate it in your next message |
| Nothing changed at all | Soft wording like "try to avoid" | Rewrite it as "Do NOT" |
That second row surprises people. Stacking rules feels like being thorough, but past a point the model spends its effort satisfying restrictions instead of doing the work, and you get a cautious, thin answer that technically breaks none of them.
✕ Over-fenced
"Do not use bullets. Do not use headings. Do not exceed 300 words. Do not use jargon. Do not speculate. Do not repeat. Do not use passive voice. Do not open with context. Do not use adjectives."
✓ Trimmed to four
"Write 300 words as flowing paragraphs, opening with the finding. Do NOT use bullets, headings, passive voice, or any figure absent from the file."
When a rule fails, resist the urge to start the whole prompt again. Change one element, rerun, and see what moves — the diagnostic approach in our guide to refining AI output without starting over applies directly here.
Start with one rule. Take the single edit you make most often, write its path, write its fence, and put it in your account settings this afternoon. That one pairing does more than any list of clever phrasings, because negative constraints in prompting only ever earn their keep one way — fewer edits on the first draft, every time.