Keep your buyer personas in an icp.md file in the repo, then have Claude Code run a background workflow of subagents. Six agents, one per persona, read the page cold and list every line where they would bounce. Three more rewrite the hero, the offer and the CTA, working only from those objections and never from each other. Three blind judges then score the original and every rewrite against the objections, and the winner comes back as one diff. Save the run as a workflow so anyone on the repo can run it as a slash command. Optionally feed it past winning headlines, support tickets and reviews. The judges never see who wrote what, so you get less self-grading, but you still need a real A/B test before calling a winner.
Prompt
tear down the landing page at [URL] 1. six agents, one per buyer persona in ./icp.md. each reads the page cold and lists every line where they'd bounce 2. three agents rewrite the hero, the offer and the CTA from those objections only, never from each other 3. three blind judges score the original and every rewrite against the objections 4. return the winner as one diff