A feature ships Wednesday and Product Marketing needs a website hero, a founder LinkedIn post, a segmented customer email, an in-app banner, a changelog entry, and an X thread. Each takes a few hours, each goes to a different reviewer with different taste. By the time everything is approved, the feature has been live two weeks and the launch moment is gone. Or PMM cuts corners, ships the LinkedIn post, skips the email, and enterprise customers never hear about the feature built for them. The Launch Comms Agent replaces the week of writing with 4 minutes of generation.
Seven moves, in order
A webhook fires when a production deploy succeeds, the release has non-empty release notes, and a Linear ticket has moved to Launched. A Thursday morning sweep catches features that shipped quietly. The agent loads the context: the Linear ticket, the Notion doc the PM wrote when the prototype shipped, the PRD, the release notes, the original customer story, and the final UI. Richer context makes better drafts.
It determines the audience per channel (LinkedIn is for PMs and ops leaders, the email is segmented by tier, the banner is only for eligible accounts, the changelog is for everyone) so the drafts do not all sound like one press release. It loads the voice.yaml rules, then runs six concurrent generations, one per channel, each with the full context plus a channel-specific prompt. Everything lands in the PMM review queue tagged pending review, and once a channel is approved the agent records it so the next run learns from what PMM actually published.
Why this works where past AI copy failed
Three design choices, each fixing a thing that killed earlier attempts. It has the full context, not just a prompt: the difference between "write a LinkedIn post about our feature" and a post grounded in the specific feature, the customer story, the real numbers, banned phrases, and a linked approved example is the data layer and voice config, not the model. It respects the review step: previous tools tried to autopost, the first 20 posts were embarrassing, and team trust died. This one always tags drafts pending and never sends. And it closes the loop on edits: every edit PMM makes becomes training signal, so after a quarter the drafts arrive much closer to publishable.
Start with three channels
Write a voice.yaml for your product: banned phrases, the tone your founder actually uses, two approved LinkedIn posts as positive examples. Pick one feature shipping in the next two weeks and hand-load its context into a Claude Code project. Ask the agent for three channels only: LinkedIn, customer email, changelog. Review with PMM, capture what they edited as new voice rules, then expand to six on the next launch. Within a quarter, the first draft on every channel arrives four minutes after deploy, and PMM's job shifts from writing to sharpening.