How Rally is used in this channel's material Rally appears as the concrete product built on the idea of simulating audiences with language models. Michael Taylor traces its origin: the entry point is roleplay — because model training data is so wide, a model can play almost any character, and the logical leap is if it can role-play one character, why not many? Turning one persona into a whole batch turns a single prompt into a focus group, an audience you test ideas against in a risk-free environment before shipping anything to the real world. [^01a0f15a-e590-7f2d-8b7d-8afcf1ca33f3@244] Rally grew out of a column post called “Personas of Thought” — a prompt that generated relevant personas, queried each one, and filled in their answers, which Taylor then scripted to automate. The first persuasive test was an internal debate at Every over a “meta media company” vs. “multimodal media company” description — and the simulated crowd made a CEO change his mind, because each persona gives its reasons, not just a verdict, and unlike real feedback loops (X, sales calls, headlines), you can rewind the clock and re-ask “what if I'd said it differently?” as a fair A/B test. The demo shows the actual flow at askrally.com: pick an audience (Hacker News, Millennials, dog lovers in Dallas), ask it a question or A/B test email subject lines, then read the results three ways — the voting tally, each individual persona's stated reasoning, and a summary of the crowd's rationale. [^01a0f15a-e590-7f2d-8b7d-8afcf1ca33f3@2829] The channel holds the tool honestly: results depend heavily on which audience you spin up and which question you ask, and Taylor's own framing is “Disclaimer — this could be wrong. It's just another piece of information.” Rally is used to sharpen questions and instincts; validation still comes from real tests. The grounding for why it works: a two-hour interview transcript yields a simulated persona ~80% accurate on traits and decisions — roughly as consistent as the real person is with their own past self. [^01a0f15a-e590-7f2d-8b7d-8afcf1ca33f3@408]