propose
Propose — research an idea, then propose what to build
The owner has an idea, or a product in use needs its next step. The agent reads what the owner shared, looks for what is missing, sets the idea against that evidence and proposes something the owner reviews, the interface first. What comes out is informed hypotheses: until people use the product and pay for it, every claim is a hypothesis with a source, never a finding.
This is not an interview. Questions come last, and only the ones the evidence cannot decide. An interview sharpens what the owner already thinks; this brings in what the owner doesn't know yet.
Not for an agreed change, a bug or a small UI change: those go straight to a plan or a build. Not when the owner has already decided and asks for the build.
Who decides. The owner owns the thesis, the segment, what gets built and when. The agent owns the evidence, the proposal and the prototype, and says plainly where the evidence and the owner's idea disagree.
Steps
1. Read what was shared, then what already exists
Everything the owner gave: notes, research, data, links, screenshots, competitor names. Then the product's own record, if there is one: the brief, the decision log, an earlier proposal. Use the repository's existing equivalent of a proposal before creating one; a second document that restates the first is how decisions get lost between them. On a re-run, start from the existing proposal (§ Running it again).