Closer to the question.
The less meaning people or AI have to reconstruct before reasoning begins, the less time, context and uncertainty the task requires.
You ask in the words the work arrived in — "what could break if subscription cancellation changes" — and get back the parts of the product that answer it. Because the match runs against descriptions of what the product does, it returns things that share no word with your question at all. Nobody has to translate the question into file names or search terms first.
Ask in the wording the task was assigned in, not in search terms.
Read what came back: areas, records, actions, endpoints, screens and the customer paths that cross them.
Follow one result you did not expect and check it against the code.
Give the same question to your agent over MCP and compare what it starts from.
This is not a search over your repository, and it does not pretend to be one: a question that names nothing the product does gets an honest "nothing matches" rather than the nearest text hit. Each question is also scored on its own — the workspace sends the question and the repository and nothing else, so it does not build on what you asked a moment ago. The worked example published on the site is measured on Supabase; the wording of the question is ours, only the answer is the product's.
See it on a real product.
Three models are published in full and open without an account. Ask one of them the question this page is about.