What a prompt has to carry
A language model writes the most plausible continuation of what it is given, as how language models generate text explains. If the prompt is generic, the plausible continuation is generic. The fix is not cleverer wording; it is a fuller brief. Six elements cover almost every marketing task.
| Element | What it answers | What happens when it is missing |
|---|---|---|
| Audience | Who will read or use this, what they know and what they want | Copy written for nobody in particular, at the wrong level |
| Task | The single job, stated as an action and an output | The model answers a broader or different question |
| Material | The facts, sources, data or draft to work from | Invented specifics that sound right |
| Constraints | Length, voice, claims allowed, words to avoid, legal lines | Off-brand tone and claims the business cannot support |
| Format | The shape of the output: table, list, headings, fields | Prose that has to be reworked before use |
| Quality bar | What good looks like and how it will be judged | Output that is fine in general and wrong for the purpose |
The same list appears, in other words, in any good content brief, and for a single content piece the AI-assisted content with review guide shows how the brief carries through drafting and editing; this guide stays with the prompt structures that work across tasks. The content briefs guide covers audience, angle, proof and voice for a piece of content, and the creative brief template does the same for campaign work. A prompt can often start by pasting the relevant parts of an approved brief, which also keeps the work tied to the team's content strategy rather than to whatever the person prompting had in mind that day.
Two habits make every pattern below work better. Separate instructions from material clearly, for example with headings such as "Instructions" and "Source material", so the model does not treat a pasted competitor page as something to follow. And tell the model what to do when the material does not cover something: "write 'not in source' rather than filling the gap" turns an invisible invention into a visible question.
The brief prompt
The brief prompt is the default pattern for producing something new: a draft, an outline, a set of headlines, an email. It is simply the six elements in order.
Audience: [who, what they already know, what they are trying to do]
Task: [one action and one output, e.g. "Write the opening section of a buying guide"]
Material: [pasted facts, approved claims, product details; use only these]
Constraints: [length, voice notes, claims not allowed, words to avoid]
Format: [headings, bullet limits, fields to return]
Quality bar: [what a strong answer does; what would make it unusable]
If the material does not cover something, write "not in source".
The audience line does the most work and is the most often skipped. "Small business owners" is weak; "owners of independent cafés comparing card readers for the first time, worried about fees and contracts" gives the model something to write to. A buyer persona, if the team has one, belongs here in a sentence or two. The task line should name one job. "Write a landing page and some ads and a few emails" is three tasks with three audiences in different states of mind, and the model will average them.
Examples first
Some qualities are easier to show than to describe: a brand voice, a house style for product titles, the level of detail in a good meta description. The examples-first variant adds two or three approved examples to the brief prompt and asks the model to match their qualities without copying their content. Choose examples that differ from each other in subject, so the model picks up the pattern rather than the topic, and say what the examples share: "short sentences, no exclamation marks, the benefit before the feature".
Extraction and structuring
Many marketing tasks start with a pile of text that needs to become data: reviews to be tagged, a competitor's pricing page to be compared, interview notes to be coded, a product specification to be turned into fields. The extraction pattern asks for a table or a fixed set of fields, never prose.
From the material below, return a table with one row per [item] and these columns:
[first field], [second field], [third field], exact quote, location in source.
Use only what the material states. Leave a cell blank if it is not stated.
Do not merge items or summarise.
Three instructions do the heavy lifting. Fixed columns stop the model inventing categories as it goes. "Exact quote" and "location in source" make every row checkable by search. And "leave blank if not stated" stops it filling cells with what is probably true. Extraction is usually the first step of a longer job: once a person has checked the table, it becomes the material for a brief prompt, and the facts in the draft can be traced back to it. The AI for research synthesis guide applies this pattern at scale to interviews and open survey answers.
Critique and revise
A model asked to improve a draft with no criteria will change it, but not necessarily for the better. The critique and revise pattern separates judging from rewriting and makes the criteria explicit.
First, ask for a critique against named criteria, returned as a list of issues with the passage each refers to: does the opening state the reader's problem; does every claim have support in the material; does the CTAThe prompt that tells a reader the one action to take next, such as a button or link. match the stage the reader is at; is anything here not allowed by the constraints. Second, a person reads the critique, discards what is wrong and adds what was missed. Third, ask for a revision that addresses only the accepted points and leaves the rest of the draft alone.
The pause in the middle is the point. Models critique confidently and sometimes wrongly, and a revision that acts on every suggestion can lose what was good about the draft. Critique against criteria is also a useful review aid for drafts written by people, as long as the reviewer stays in charge of what changes.
Generate, then rank against criteria
Variation is cheap with a model; judgement is not. The generate-then-rank pattern uses the cheap part to widen the options and keeps the judgement explicit.
Ask for a set of distinct options, such as headlines, subject lines, angles or offers, each taking a different approach rather than rewording the same idea. Then, in a separate step, ask the model to score each against criteria you supply, with a reason for each score: clarity for the named audience, fit with the search intent of the target query, Message MatchHow closely a landing page repeats the promise and wording of the ad or link that sent the visitor. with the ad or email that leads to the page, compliance with the claims rules. Finally, a person picks.
Ranking is a filter, not a verdict. The model's scores show which options meet the written criteria; they do not predict how customers will respond, which only a test with real traffic can show. Use the ranking to cut a long list to a short one, then put the short list through the team's usual review and, where the volume justifies it, an A/B test.
Rewriting for a reader and a channel
Much marketing work is adaptation: a technical document rewritten for buyers, a long article cut into an email, a UK page localised for another market, a webinar turned into social posts. The rewriting pattern names both ends of the change.
Source: [the approved text]
Original reader: [who it was written for]
New reader: [who it is for now, what they know, what they need from it]
Channel: [where it will appear and its limits: length, format, conventions]
Keep exactly: [facts, figures, legal wording, product names]
May change: [order, examples, tone, level of detail]
The "keep exactly" line is what makes this pattern safe. Rewriting is where facts drift: a qualification is dropped to save words, a figure is rounded, an example is swapped for one that does not apply in the new market. Listing what may not change, and checking those items afterwards, catches most of it. The AI for content guide covers repurposing and localisation as content stages and the review each needs.
A cookware brand rebuilds one prompt with the six elements
Saxonmere Cookware, an invented maker of cast-iron pans, has been using the prompt "Write a product page for our 26cm skillet" and getting copy that calls the pan "non-stick" and "dishwasher safe", neither of which is true. The content lead rebuilds it as a brief prompt. Audience: home cooks who already own non-stick pans and are considering cast iron for the first time. Task: write the description section of the page, 150 to 200 words. Material: the specification sheet and the 4 approved claims, pasted in. Constraints: never describe the pan as non-stick; care instructions must say hand-wash only. Format: 1 short paragraph and 4 bullets. Quality bar: a first-time buyer should understand seasoning in one read. Of 10 drafts from the old prompt, 7 contained a claim not in the specification; of 10 from the new one, 1 did, and the reviewer caught it against the pasted sheet.
Patterns by task
The patterns combine. A typical content job runs extraction on the sources, a brief prompt on the extracted facts, critique and revise on the draft, and generate-then-rank on the headlines.
| Task | Main pattern | Supporting pattern | What the prompt must include |
|---|---|---|---|
| Blog or landing page draft | Brief prompt | Critique and revise | Audience, angle, approved facts, voice, "not in source" rule |
| Ad headlines and descriptions | Generate, then rank | Examples first | Character limits, claims rules, landing page message |
| Email subject lines | Generate, then rank | Rewriting for a reader | Segment, offer, what the email actually contains |
| Review and feedback analysis | Extraction and structuring | Critique and revise | Fixed tags, exact quotes, source references |
| Competitor page comparison | Extraction and structuring | Brief prompt | The fields to compare, blank-if-missing rule |
| Report narrative | Brief prompt | Extraction and structuring | The figures, their definitions, who reads the report |
| Product descriptions at volume | Examples first | Extraction and structuring | Specification fields, house style examples, banned claims |
| Localising or repurposing | Rewriting for a reader and a channel | Critique and revise | Keep-exactly list, channel limits, new reader |
Once a prompt works for a recurring task, it belongs in a shared, tested library rather than in one person's chat history; building a team prompt library shows how. Whether it works well enough to rely on is a question for the method in testing prompts and models.
Common mistakes
- Briefing the tool less than you would brief a freelancer. If a capable person would need to ask questions, the model needs the answers in the prompt.
- Leaving out the material. Without facts to draw on, the model supplies typical ones. Paste what must be true and say to use only that.
- Mixing instructions and material. Pasted text with instructions inside it can be followed as if it were the brief. Label the sections.
- Asking for several jobs at once. One prompt, one task, one audience. Chain steps instead.
- Accepting every critique. A model's critique is a list of suggestions; a person decides which are right.
- Treating a model's ranking as a forecast. Scores against written criteria are a filter, not a prediction of customer response.
- Collecting magic phrases. Phrasing tricks vary by model and fade with updates; the six elements do not.
- Rewriting without a keep-exactly list. Facts, figures and legal wording drift quietly in adaptation.
