How to Prompt AI to Write a Press Release
Most AI press release prompts just ask for facts and hope. Here is the actual structural contract to feed the model, plus a technique for fixing fake-sounding executive quotes.
Prompting AI to write a press release fails for one predictable reason: the prompt hands over facts and asks for "a press release" without specifying the format's actual rules. A press release has a fixed contract: headline, dateline, a lead paragraph answering who, what, when, where and why in one paragraph, supporting detail in descending order of importance, a quote block, and a boilerplate paragraph. Skip that contract in the prompt and the model defaults to generic PR-speak: stacked superlatives, vague numbers, and a quote that could have come from anyone at any company. The fix is not a better topic description. It is a better structural spec.
Why AI-written press releases read fake
Ask most AI tools for a press release and you get the same tells every time. Adjectives do the work nouns and verbs should be doing: unprecedented, cutting-edge, next-generation. Numbers get hedged into vagueness: significantly increased, substantially improved, instead of an actual figure. The executive quote reads like it was assembled from a template, because functionally it was.
None of this is a model capability problem, it is a specification problem. The model has read thousands of real press releases mixed with thousands of imitation ones, and without a structural anchor it drifts toward the average of both: safe, vague, forgettable.
The structural contract of a real press release
A press release a journalist will actually use follows the inverted pyramid: the most newsworthy information comes first, and each paragraph after that matters a little less than the one before it. That shape has specific parts, and each one has a job.
Section | Job | What breaks when AI improvises it |
|---|---|---|
Headline | States the news in one active-verb sentence | Turns into a slogan instead of a claim |
Dateline | CITY, State, Month Day, Year, anchors the release in time and place | Gets dropped entirely |
Lead paragraph | Answers who, what, when, where, why in about 40-60 words | Becomes three vague sentences instead of one dense one |
Body paragraphs | Supporting detail, most important first | Facts scatter out of priority order |
Quote block | A named person's perspective, not a restated fact | Reads like marketing copy with a name attached |
Boilerplate | Standard company description, consistent release to release | Gets rewritten each time, drifting into ad copy |
Most people never tell the model any of this. They describe the news and trust the label "press release" to carry the format with it. It does not, models pattern-match to a label loosely; you have to supply the skeleton yourself.
Feed the model the structure, not just the facts
Give the model the section labels, a word count target for each one, and the order, before you give it the facts. Treat it like briefing a new writer on their first press release, not like asking a friend to summarize some news.
This is the same discipline covered in prompt engineering fundamentals: the more of the output's shape you specify up front, the less the model has to guess, and guessing is where hollow language comes from.
You are drafting a press release. Follow this exact structure. Do not add adjectives not directly supported by the facts below.
STRUCTURE:
1. HEADLINE: One sentence, active verb, states the news directly. No words like "revolutionary," "unprecedented," or "next-generation."
2. DATELINE: [CITY, STATE] - [Month Day, Year]
3. LEAD PARAGRAPH (40-60 words): Answer who, what, when, where, why in a single paragraph, not split across sentences.
4. BODY (2-3 paragraphs): Supporting detail, most important first. One claim per paragraph.
5. QUOTE BLOCK: Give me THREE quote options from [name, title], each a different angle:
- Option A, customer-benefit framing: what changes for the person using this
- Option B, market-context framing: why this matters given the current situation in the space
- Option C, forward-looking framing: what this enables next, a concrete step
Each quote: 1-2 sentences, no restated facts from the lead.
6. BOILERPLATE: 2-3 sentence standard company description, factual, no superlatives.
RULES:
- No words like "significantly" or "substantially" without a number attached.
- No passive-voice hedging, such as "it is expected that."
- Every claim must trace to a fact below. Do not invent numbers, dates, or outcomes.
FACTS:
[Company name, what happened, date, key numbers, names and titles of people to quote, why this matters to the audience.]The quote problem: draft three options, not one
Executive quotes are where AI-written releases fall apart fastest, because a generic quote is worse than no quote. Readers can tell when a quote just restates the headline in a different font.
The fix is to stop asking for one quote and ask for three, each on a different angle:
Customer-benefit framing: what changes for the person using the product or service
Market-context framing: why this matters given where the category currently stands
Forward-looking framing: what this makes possible next, stated as a specific step
Generate all three, then a human picks one, edits it, or blends two. Never ship the model's first quote unedited; voice is the entire point of a quote. If you already maintain a documented voice reference, reuse it here, the same techniques in how to prompt AI to match your brand voice apply directly to quotes. For the market-context option, pull the sharpest line from work already done; a competitive analysis is often the fastest source. And for a quote that needs a specific voice rather than a generic spokesperson, the persona techniques in how to get AI to answer like a specific expert carry over well.
A worked example
Fictional facts: Lumen Ops, a scheduling platform for multi-location restaurant operators, is launching automated shift scheduling built from sales forecasts and staff availability.
Headline: Lumen Ops Launches Automated Shift Scheduling for Restaurant Chains
Dateline: AUSTIN, TX - March 4, 2026
Lead: Lumen Ops, a workforce management platform for multi-location restaurant operators, today announced automated shift scheduling, which builds weekly schedules in under two minutes. The feature is available now to all customers running three or more locations.
Quote options generated for review:
Option A (customer-benefit): "Scheduling used to take a manager four hours every Sunday. Now it takes about four minutes." - Dana Ruiz, Operations Director, a Lumen Ops customer
Option B (market-context): "Labor costs at restaurant chains kept climbing faster than menu prices again this year, and most operators still schedule with spreadsheets." - CEO, Lumen Ops
Option C (forward-looking): "Scheduling was the first piece. The same forecasting engine is what we are using for inventory ordering next quarter." - CEO, Lumen Ops
An editor picks Option A because it pairs with the efficiency claim in the lead, then adds a specific number in place of "about four minutes" once one exists. That single edit is the difference between a quote that sounds like a person and one that sounds like output, and the same discipline is what makes a launch worth turning into a case study later.
The checklist: what makes a release read like AI wrote it
Superlative stacking: groundbreaking, unprecedented, next-generation, describing something the facts alone do not support. If removing the adjective kills the sentence, it was not carrying news.
Vague quantification: significantly, substantially, considerably, with no number attached. Replace with the real figure or delete the claim.
Passive-voice hedging: phrasing like "it is anticipated that," which removes an actor to avoid committing. Name who is doing what.
Uniform sentence rhythm: every sentence the same length, no short line carrying the actual news. Reads like one unedited pass.
A quote that repeats the lead paragraph in different words. If deleting it loses no information, it is padding with an attribution.
If a draft still reads generic after you have supplied the full structure above, the problem is still in the prompt, see how to fix a bad AI prompt before you touch the draft by hand.
FAQ
Can AI write a full press release from scratch? It can produce a structurally correct first draft once you supply the section labels, word counts, and facts. Treat that draft as a starting point, not a finished release; quotes and soft claims still need a human edit before it goes out.
What is the biggest mistake people make when prompting AI for a press release? Describing the news and trusting the label "press release" to carry the format. You have to specify the headline, dateline, lead paragraph, quote block, and boilerplate explicitly, or the model defaults to generic marketing language.
How do you make an AI-generated quote sound like a real person said it? Ask for three quote options at different angles, customer-benefit, market-context, and forward-looking, then have a human pick one and edit it. Never publish the model's first quote attempt unedited.
What words should you cut from an AI press release draft? Unsupported superlatives, vague quantifiers like significantly or substantially with no number attached, and passive-voice hedges that avoid naming who did what.
How did this land?
About the author
Developer Advocate
Steve builds something with Swarmz every week and writes up what worked, what broke, and what he'd do differently. Tutorials and hands-on guides are his lane.
