Negative Prompting and Constraints: Tell AI What to Skip

AI loves clichés. It loves the words 'delve,' 'tapestry,' and 'in conclusion.' If you ban them up front, you save yourself a rewrite. Negative prompting is just listing what to avoid.

Some examples

• 'Write a poem about autumn. DO NOT use the words leaves, crisp, or pumpkin.'
• 'Summarize this article. No bullet points. No headers. Just one paragraph.'
• 'Help me brainstorm. Don't suggest anything I'd find on the first page of Google.'
• 'Edit this essay. Do not change my voice or add new arguments.'

Try it!

Ask AI to write a 3-sentence intro to a story you're working on. Then re-ask with 'Do not start with the weather, do not name the character yet, and do not use the word suddenly.' Compare.

AI has a 'context window' — basically short-term memory. Past a certain length of conversation, it starts forgetting your earlier instructions. If you've ever had AI 'go off the rails' deep in a chat, that's why.

Some examples

• Long chats often forget the rules you set at the start.
• Pasting a 50-page document might mean only the first part actually 'sticks.'
• Re-pasting your instructions every few turns keeps them fresh.
• Starting a new chat is sometimes the fix when AI gets confused.

Try it!

Have a long back-and-forth with AI on one topic, then test if it remembers a detail from your very first message. If it doesn't, you've hit the context limit.

AI loves to over-explain, add disclaimers, and start every answer with 'Great question!' If you list anti-patterns explicitly, you save tons of editing time.

Some examples

• 'Do not start with a compliment.'
• 'No bullet points unless I ask for them.'
• 'Never say "As an AI, I cannot..." — just answer.'
• 'Do not add a summary at the end. End at the last point.'

Try it!

Make a personal 'banned phrases' list. Paste it at the top of every prompt for a week and notice how much cleaner the answers feel.

You can tell Claude or ChatGPT 'don't apologize at the start' or 'no emojis' and it'll mostly comply. But 'don't' instructions are weaker than 'do' instructions because the model still has to think about the forbidden thing. 'Be direct' beats 'don't apologize' even though they aim at the same target.

Some examples

• Weak: 'Don't write more than 100 words.' Strong: 'Write 80–100 words.'
• Weak: 'No emojis.' Strong: 'Use plain text only.'
• Weak: 'Don't be cringey.' Strong: 'Match this tone: [paste sample].'
• Weak: 'Don't make it generic.' Strong: 'Reference my hobby (basketball) at least twice.'

Try it!

Find a prompt you use that's full of 'don't.' Rewrite each 'don't X' as 'do Y instead.' Test both versions.

Negative prompting is listing what you want the model to avoid. It's especially effective for the AI's default tics — flowery language, em-dashes everywhere, 'I hope this helps!' sign-offs. The bans are short, memorable, and cheap.

Some examples

• 'Don't apologize, don't hedge, don't add disclaimers.'
• 'No bullet points — write in flowing paragraphs only.'
• 'Avoid the words: delve, leverage, robust, comprehensive.'
• 'Do not invent facts. If you don't know, say so.'

Try it!

Make a personal 5-item ban list of AI tics that bug you. Paste it at the top of your next 3 prompts. Notice the cleanup.

Big context windows tempt you to dump everything. But model attention degrades as the window fills — important stuff gets buried. Treat the context like a budget: include what's needed, summarize the rest.

Some examples

• You summarize old chat turns instead of forwarding the whole transcript and Claude stays sharp.
• Cursor's 'include only relevant files' beats 'include the whole repo' on accuracy.
• ChatGPT with the top 5 relevant docs outperforms the same prompt with all 50 stuffed in.
• Trimming a 50-page PDF to the 3 relevant pages improves recall by a noticeable margin.

Try it!

Take a prompt where you usually paste a lot. Cut it in half by removing the least-relevant parts. Compare.

Asking for 'a story' gives generic. Asking for 'a 50-word story in second person about a vending machine' gives memorable.

Some examples

• Add a length constraint.
• Add a perspective or voice constraint.
• Add a format constraint (haiku, listicle, dialog).

Try it!

Take any boring prompt. Add 3 constraints (length, voice, format). Compare.

Understanding "Constraints unlock creativity (counterintuitively)" in practice: Prompting is a skill: the more specific and structured your input, the more useful the output. Tight constraints in your prompt make AI output more creative, not less — and knowing how to apply this gives you a concrete advantage.

• Use role, context, task, and format in every prompt
• Iterate: treat first outputs as drafts, not finals
• Use few-shot examples for complex formatting tasks
• Test prompts at different temperatures for creative vs. factual tasks

1. 1Rewrite one of your best prompts using role + context + task + format
2. 2Ask an AI to critique your prompt and suggest improvements
3. 3Compare outputs from two models using the same prompt