I have watched too many product decisions turn into archaeology. Someone asks why a feature exists, and the answer is a Slack thread from two years ago, a person who has left, and a conclusion everyone repeats without remembering the argument behind it.
An agent makes that worse in a specific way. It will produce a confident market analysis in which the customer quotes are indistinguishable from its own inventions, and the analysis is more polished than anything a tired human would write on a Friday. Confidence is not evidence, and a model has no internal penalty for inventing one.
This skill is the opposite habit, and it exists because the two failure modes it prevents cost different kinds of money.
The idea
Knowledge is a deliverable, with a date and a status.
- A claim carries its source, or it is labelled a hypothesis with a confidence level. “Customers want X” is either an interview quote with a name and a date, or it is a hypothesis, and the document says which.
- A ruling is written down with the evidence that produced it and who ruled. Six months later the question “why did we do this” has an answer that is not a memory.
- Options stay open until something eliminates them. Not until someone senior gets tired of the discussion, and not because a roadmap template had room for one.
- Deferrals are written down too, with the trigger that would reopen them. A rejected idea with no record comes back every eighteen months, re-litigated from scratch.
That last part changes behaviour most. Most plans are not wrong at the start; they are wrong because an early assumption quietly became a decision, and by the time it is expensive nobody wants to reopen it.
What it works against
The roadmap that is really a set of opinions, and the meeting where a decision is “made” and then re-made every three weeks by whoever remembers it most confidently.
It also works against the agent failure it was written for: a competent writer with no memory of what it invented, generating a document that reads like research.
What it is not
Not a market research service, and not a substitute for talking to customers. It organises evidence; it does not produce any. If the only sources are the model’s imagination and my own memory, the output says so, and that is the point.
Available on request; the skill text is not published here.