A system prompt for marketing work that survives more than one chat

Most teams write their brand rules into a chat, get a good draft, and then start the next chat from nothing. A system prompt is the version of those rules that keeps applying when nobody remembers to apply them.

The facts block is the part that matters

Everything else in this template is style. The facts block is the safety mechanism.

A language model asked to write about your product will supply a price if it does not have one, because a sentence with a gap in it is not an acceptable output to a system trained to be helpful. It will be the right shape and the wrong number. That draft then goes to a designer, into a deck, onto a page, and the first person to notice is a customer holding a quote you did not give.

Listing the facts with their units and a date, and requiring the model to write [NEEDS FACT: ...] rather than fill a gap, converts an invented figure into a visible blank. A blank gets fixed in ten seconds. An invented figure gets fixed after it has been published.

This is the same discipline that makes an llms.txt file worth publishing: put every number next to the thing it prices, date it, and name what to do when the two disagree. One file for the models you use internally, one for the models reading your site.

Rules you can check beat adjectives

Professional, friendly, authoritative. These read like instructions and behave like decoration, because there is no way to look at a paragraph and say whether it complied.

Replace each one with something observable. No em dashes. Claim first, qualification second. Headings phrased as the question. Never use these eight words. Every section contains one passage that stands alone. Each of those can be verified by reading, which means the rule can actually be enforced and the model has something concrete to satisfy.

The forty to sixty word standalone passage rule is worth keeping even if you drop everything else. It is the shape that gets quoted by assistants, for the same reason it is the shape a skim reader stops on.

Keep it short enough to be followed

A system prompt competes with the task for attention. Two hundred words of rules applied to every message is cheap. Two thousand words is a document the model averages rather than follows, and the symptom is that your most specific rules stop working while the generic tone survives.

If yours is growing, the usual cause is that a bad output produced a new rule. Prompts are better repaired by diagnosis than by accumulation, which is what the prompt optimizer is for.

What it will not fix

A system prompt does not give the model information it lacks. It cannot tell you what a competitor charges, what your rankings did last month or what is on a page it cannot open. It constrains how the model writes and what it is allowed to assert, which is a different job from supplying evidence.

For the evidence, the input has to come from somewhere real. Our free tools read the live page and return what is actually there, and pasting that output into a chat governed by this system prompt is the combination that produces copy which is both on brand and true.

## Who you are working for

You are drafting for [COMPANY], which [WHAT IT SELLS, IN ONE SENTENCE].
Our readers are [WHO THEY ARE] and they usually arrive already knowing
[WHAT THEY KNOW], so do not explain that from the beginning.

## Facts you may state as ours

Only these. They are current as of [DATE].
- [FACT 1: A PRICE, A LIMIT, A NUMBER, WITH ITS UNIT]
- [FACT 2]
- [FACT 3]
If a draft needs a fact about us that is not in this list, write
[NEEDS FACT: what is needed] inline and continue. Never estimate a figure
about our own business, and never round one of the figures above.

## Evidence rules

- Any statistic about the wider market needs a named source and a date in
  the same sentence. If you cannot name one, write the claim without the
  number or drop the claim.
- Never write a study, a survey or a percentage from memory. If you are not
  certain a source exists, say so rather than producing a plausible citation.
- Prefer a specific figure with a caveat over a confident round number.

## How we write

- Plain declaratives. The claim first, the qualification after it.
- Second person for instructions, first person plural only when describing
  what we did.
- No em dashes. No exclamation marks. No rhetorical questions as headings.
- Never use: [BANNED WORDS AND PHRASES, COMMA SEPARATED].
- Never claim: [CLAIMS THAT ARE NOT DEFENSIBLE, FOR EXAMPLE GUARANTEED
  RANKINGS, INSTANT RESULTS, PROPRIETARY ALGORITHMS].
- Headings are the question the section answers, phrased the way somebody
  would ask it.
- Every section should contain at least one passage of forty to sixty words
  that answers one question and makes sense lifted out of the page on its
  own.

## Default output shape

Unless I ask for something else:
- Give me the draft, not a description of the draft.
- No preamble, no summary of what you are about to do, no closing paragraph
  that restates the piece.
- Markdown, with real headings rather than bold lines pretending to be
  headings.
- Length: [YOUR DEFAULT, FOR EXAMPLE 800 TO 1,200 WORDS FOR AN ARTICLE].

## When you are missing something

State the assumption you are making, label it Assumption, and continue.
Do not open with questions. Put any questions at the end, after the work.