Why AI breaks your rules
A model writes its answer one token at a time, a token being a word or a piece of one, and it keeps no tally as it goes. A limit of 150 words is a target it aims for, not a ceiling it can see, so the answer lands near it and sometimes well past it.
Other rules break for other reasons. Models are trained to be friendly, so they open with a warm line and close with an offer of more. A format rule at the top of a long prompt ends up far from where the answer starts. A rule written only as what not to do gives the model nothing to do instead. And when two rules collide, the model quietly drops one. Each kind of rule has a wording that holds better.
| Kind of rule | Why it breaks | State it like this |
|---|---|---|
| A word limit | No running count, so the limit is a guess. | “2 paragraphs of 2 or 3 sentences each”, with the reason, then count. |
| How many things | An unnumbered list has no visible end. | “4 subject lines, numbered 1 to 4.” |
| A format, such as a table or JSON | Stated early, then buried under the material. | Last, after the material, with a short example of the form. |
| Banned words or habits | Says what to leave out, not what to write. | What to write instead; the banned list as a final check. |
| Only the answer | Trained to open and close politely. | “Start with the subject line. Write nothing after the sign-off.” |
| Two rules at once | One has to give, and the model picks which. | “If it cannot fit, keep the dates and cut the background.” |
Write the lead item for our customer newsletter about the new invoice reminders, from these notes: {{feature_notes}}. 1Use only the facts in the notes, and write [CHECK] where one is missing. Write 2 short paragraphs of 2 or 3 sentences each, because the newsletter shows about 120 words. 2Then give 4 subject lines, numbered 1 to 4. 3Where you would call the feature exciting, say what it does instead. 4Start with the first paragraph and stop after the last subject line. 5Format: plain text, no headings or bold, laid out like this example: “You can now send reminders automatically…” and then “1. Stop chasing late invoices”.6
- 1Material firstLong notes go above the rules, so the rules are the last thing it reads.
- 2A limit it can seeParagraphs and sentences, with the reason for the limit.
- 3A numbered countEach number it writes shows how far it has come.
- 4What to write insteadA plain instruction in place of a ban.
- 5Only the answer, definedWhere to start and where to stop.
- 6Format last, with an exampleThe last rule it reads, and a form to copy.
1. Give limits in sentences or items
Ask for a shape instead of a word count. “2 paragraphs of 2 or 3 sentences” or “5 bullet points of one line each” is something the model can follow as it writes, because each paragraph and bullet is a unit it has finished or not.
When the limit is hard, such as a form field or a print column, aim below it and give the reason: “Under 120 words, because the form cuts off at 150.” Anthropic’s prompting best practices show that a reason helps the model apply a rule more widely than its wording.
Then count. If the answer runs over, ask for the cut in a follow-up and say what survives it: “Cut this to 120 words. Keep the dates and the price, and cut the background first.” The cut-to-a-limit prompt does the same in one step.
Keep it under 150 words.
Not quite. A word count is a target the model cannot see while it writes, so the answer can land well past it.
IMPORTANT: NEVER go over 150 words.
Not quite. Capitals add urgency, not a way to count. Anthropic advises plain wording for its recent models, which already pay close attention to instructions.
Write 2 paragraphs of 2 or 3 sentences each.
Yes. Paragraphs and sentences are units the model can see as it writes them, so the shape holds where a word count drifts.
Be concise.
Not quite. Concise means something different to every reader, and to the model. Give it a shape to fill.
2. Number every count
“Some subject lines” leaves the number to the model. Write the count as numbers it fills in: “4 subject lines, numbered 1 to 4.” Each number it writes shows how far it has come. When the items have their own limit, give both: “5 bullet points, each a single line.”
When the count must be exact, such as slides in a deck, say what to do when the material does not fit: “Exactly 5. If there are more points, merge the smallest.”
3. Put the format last, with an example
Long material goes first and your instructions after it. Anthropic’s guide advises putting long documents near the top of a prompt, above the instructions, and reports that a question placed at the end can improve answers by up to 30 percent in its tests. The format is the rule most easily lost, so make it the last thing the model reads.
Then show the form. The same guide calls examples one of the most reliable ways to steer the shape and tone of an answer. Keep the example short and mark it as one, so the model copies its layout and not its content:
Format: a table with these columns, and nothing before or after it. <example> | Supplier | Price per unit | Delivery time | | Example Co | $4.20 | 5 working days | </example>
The style of the prompt rubs off too. The guide notes that removing markdown from a prompt can reduce the markdown in the answer, so if you want plain text, write the prompt in plain text.
4. Say what to write, not only what to avoid
“Don’t use buzzwords” tells the model what to leave out and nothing about what to put in. Anthropic’s guide puts the rule plainly: tell the model what to do instead of what not to do, so instead of “Do not use markdown”, ask for “smoothly flowing prose paragraphs”. For words, name the plain alternative: “Where you would call the feature exciting, say what it does.”
Keep a list of banned words for the end, as a check on the draft: on its own, a ban tends to move the model to a near synonym. The guide to AI words to avoid has plain swaps for the usual offenders.
5. Define “only the answer”
“Just the text” means something exact to you and something looser to a model trained to be polite. Say where the answer starts and where it stops: “Start with the subject line. Write nothing after the sign-off.” For data, name the first character: “Reply with the JSON object only, starting with the opening brace.”
A stray opening line is easy to delete by hand. When the answer goes straight into a spreadsheet or another program, try the rule on several answers first.
Before
4/10 A start
- Has: What to make
- Has: Who it's for
- Missing: Background facts
- Has: Length
- Missing: Format
- Has: What to avoid
- Missing: What good looks like
- Missing: An example
- Missing: If unsure
- Missing: Layout
After, with what was added marked
10/10 Excellent
- Has: What to make
- Has: Who it's for
- Has: Background facts
- Has: Length
- Has: Format
- Has: What to avoid
- Has: What good looks like
- Has: An example
- Has: If unsure
- Has: Layout
Illustrative, for a made-up invoicing app. The word limit became paragraphs with a reason, “some subject lines” became four numbered ones, the ban on buzzwords became what to write instead, and “just the text” became where to start and stop. The format comes last, with an example. The marked words are what the second version adds.
6. Say which rule wins
Some rules cannot all hold at once. “Under 100 words” and “cover every point in the report” fail together on a long report, and the model breaks one of them without saying which. Rank them: “If it will not fit in 100 words, keep the decision and the deadline, and leave out the background.”
The same goes for a rule that clashes with your material, such as five examples from notes that hold two: “If there are fewer than five, use the ones there are and write [CHECK] for the rest.”
7. Check the answer against the prompt
Read your prompt as a checklist. Every rule in it is a line the answer either kept or broke: the length, the count, the format, the words to leave out, and where it starts and stops. Then look for what the answer added. A number, a date, a name or a link that was not in your prompt or your material is where invented facts hide.
Asking the model to check its answer against each rule catches some misses, but it counts no better the second time, so check lengths and counts yourself.
For a prompt you use every week, check more than one answer: turn the rules into checks in the Prompt Tester and run them on real inputs, as the guide to testing prompts shows.
Key takeaways
- A model writes a piece of a word at a time and keeps no count, so a word limit is a target, not a ceiling.
- Give limits as paragraphs or items, with the reason, and count the answer before you use it.
- Number every count, and say what to do when the material holds more or less than that.
- Put long material first and the format last, with a short example of the form.
- Say what to write instead of only what to avoid, and say which rule wins when two collide.
- Check every answer against the rules in your prompt, and look for facts it added.
Questions
Why does ChatGPT ignore my word count?
Because it writes a token at a time, a word or part of one, with no running tally, so a word limit is a target it estimates rather than a number it can see. Ask for paragraphs or sentences instead, give the reason for the limit, and count the result. The Instruction Check counts it for you.
How do I get ChatGPT to write exactly 500 words?
You rarely get an exact count in one go. Ask for a structure that comes out near 500 words, such as five paragraphs of four or five sentences, then count. If it runs over, ask it to cut and say what to keep; if it falls short, ask it to expand one named paragraph. The cut-to-a-limit prompt does the trimming.
Why does AI keep using words I told it not to use?
A ban says what to leave out but not what to write, so a model that avoids one word can reach for a close synonym. Say what to write instead, and keep the banned list as a check on the draft. The Robot-Speak Checker marks the usual AI words and habits in any draft.
How do I make ChatGPT give only the answer, with no introduction?
Say where the answer starts and where it stops: “Start with the subject line and write nothing after the sign-off.” A model trained to be polite reads “just the answer” loosely, while a first and a last line are rules it can follow and you can check.
Does writing rules in capital letters help?
Rarely. Capitals add urgency but no information. Anthropic’s prompting guide advises softening that kind of wording for its recent models, which already follow instructions closely and can overdo a shouted rule. Give the reason instead.
Why does the AI follow my rules at first and then stop?
In a long chat, the rules from your first message sit far behind the latest one. Repeat the rules that matter in the message that needs them, or start a new chat with the full prompt. Rules you want in every chat belong in custom instructions or a project.
Prompts to use with this
- Cut a text to a word or character limitYou need a text to fit a form, abstract or post limit without losing the main point.
- Improve a prompt that gives inconsistent resultsYou need a prompt you use often rewritten so it gives reliable, well-formatted answers.
- Edit a document to match your house style guideYour organization has rules for spelling, capitals, numbers and terms, and you need a draft brought into line with every change listed for review.
- Rewrite expert text in plain languageYou have technical, legal-sounding or specialist text that ordinary readers need to understand and act on.