The Product Agent

The first diagram shows where people decide. This is what it took to keep those slots from filling themselves in.

Human judgment gets a protected slot, and the model may not fill it. Themes carry an organic confidence score derived from how much evidence supports them. They also carry a human signal score, set by a person who was in the room and thought something mattered more than the volume suggested. Ranking is the composite, so a medium-confidence theme with strong human signal outranks a high-confidence theme with none. The rule that makes it work: the model must never generate, suggest, propose, or infer a human signal flag. That input exists specifically to encode what the transcript did not capture. A model filling it in would be recording its own inference as a person’s judgment, which corrupts exactly the signal the field was built to hold.

The council probes and never approves. Before an idea becomes a project, five perspectives push on it and surface blind spots. None of them endorses, greenlights, or validates the direction. A PM leaves with sharper questions, not permission to build. The decision stays where it belongs.

Citations resolve to records, not links. Every evidence claim in a brief has to resolve to an actual row in the research or feedback store, with a theme identifier and a source URL. A shared document link is reference material for a human, never just a citation. Attribution without a link is not attribution. It is what makes an evidence-backed brief checkable rather than merely confident. Anyone can follow a claim back to the interview that produced it.

Citing a theme is not the same as addressing it. We track which customer themes have a campaign working on them. A brief citing a theme as evidence does not create that link. Only a person explicitly asserting it does. The distinction sounds pedantic until you watch a coverage map fill itself in with work nobody is actually doing.

We cut the feature that would have automated it. The first version of that coverage map had an automatic matcher pairing themes to campaigns by keyword and goal alignment. At any threshold defensible enough to trust, it fired on roughly one theme in fifteen. That was a ceiling in the shape of the data, not a tuning problem. Shipped, it would have shown leadership a coverage map that looked complete and was not. We cut it and made coverage human-written.

Conflict gets flagged, not resolved. The discovery planner scores an initiative across seven factors and recommends a research depth. When a PM’s brief budgets less time than the score calls for, the system does not silently rescope and does not overrule. It plans to the PM’s number and names the gap. The PM owns the call. The system owns telling them what they are trading.

Capture happens at the source, and there is a gate on it. Before any conversation is logged, the system asks for the recording and will not proceed to a manual write-up until a person confirms there isn’t one. A research layer built on write-ups holds one person’s interpretation. A layer built on transcripts holds what the customer said. Only one of those survives being queried a year later by someone who was not in the room.

A brief cannot ship to no one. No unit of work is final without a named release audience: internal, a named early access group, or specific named customers. General availability is not a valid answer. This is the smallest rule in the system and it closes the loop, because a named audience is what makes it possible to go back and ask how it landed.