Most AI-written press releases sound like every other AI-written press release. There's a way to use Claude that produces something distinctive in your voice — but it requires more setup than the obvious approach. Here's the workflow.
Set up a Press Release Project loaded with 5-10 of your team's strongest past releases in your actual voice. Paste in the news + 2-3 quotes from real humans. Ask for a first draft following the structure of the examples. Heavy edit. The voice comes from the examples; without them, the output is generic.
Default LLM output for press releases is averaged across millions of internet press releases — which means averaged corporate-speak. "Today, [Company] is pleased to announce..." "Industry-leading..." "Cutting-edge..." All the words PR cliché contests are made of.
Voice has to be loaded as examples, not described as adjectives. "Be confident and warm" doesn't work. Three of your best past releases in your actual voice does work.
"Draft a press release about [news]. Use the structure and voice of the example releases in your knowledge. Forbidden words and phrases: \"pleased to announce,\" \"industry-leading,\" \"cutting-edge,\" \"game-changing,\" \"revolutionize,\" \"thrilled.\" Use the actual quote provided; do not invent quotes. Keep total length under 350 words."
Add these to every press release prompt:
Forbidding the cliches forces Claude to find more specific language.
Often yes if they're unedited. The voice patterns are recognizable. Heavy editing solves it.
It can draft a quote. The executive should rewrite it in their own words before approval. Don't pretend AI-drafted quotes are the executive's actual words.
30-45 minutes for first draft, 30 minutes for editing. ~half the time of writing from scratch.
Same workflow works. The SEO press release space is already saturated; voice matters more than ever.
Use enterprise-tier AI with confirmed no-training data handling. Don't paste embargoed material into free tools.