A studio project: a site that rates US political claims against the evidence, not against who made them. We designed and built the whole thing with AI: the site, the data, the dashboards and a newsroom of agents, with a person approving every verdict.
- What we did
- Product design, Web app, AI agents and skills, Data and dashboards
The project
A studio project to show what a properly built, AI-run publishing system looks like. Designed and built with AI from the first screen to the last line of code, and built correctly: a working, data-driven product with dashboards, reporting and a publishing pipeline, not a demo.
The idea
Judge the claim, not the person who made it. Every report puts a quoted, linked claim beside one left-leaning and one right-leaning article and the primary record (BLS, CBO, Census, FEC, statutes), then gives a verdict and a confidence score from 0 to 10 with its reason. The site's own line: question everything, even this site.
The newsroom
A team of agents does the legwork. One scours the news for claims worth checking, one researches, a separate fact-checker tests every source, quote and score, a writer drafts, and another compares the findings into a report. Then it comes to a person, because a fact-checking site has to have a human making the call. Every verdict gets human review, and a “Lie” rating also needs the editor's sign-off.
The rules
Contributors can each run a category with their own agents, and every agent works to the same methodology and the same strict rules about what can go on the site. The quality comes from the rules, not from who's running them. The artwork works the same way: a skill draws the illustration for each featured story.
Built to run
Routine research runs on cheaper models, and frontier-model tokens are saved for the parts that need judgment: fact-checking and verdicts. Migrations are built in, so features can be added or removed without taking the site down. Corrections are dated and logged, never quietly edited.