Synthesize evidence into testable judgments

Code incidents before patterns

Extract setting, goal, workaround, cost and failure point with source locations. Do not count multiple quotes as multiple users. Similar words can reflect different mechanisms, such as time savings versus responsibility. AI may suggest codes, but verify them against notes.

Keep the reasoning chain

For each judgment list support, negative cases, alternatives and confidence basis. Prioritize unresolved risk rather than mention count. A frequent costly workflow may warrant study, but delivery and payment need separate evidence. Set success and failure criteria before testing.

Worked example · fictional

From a problem to inspectable work

Fictional: three users say save time, referring to submission, reconciliation and approvals.

1. Code tasks and costs separately instead of assuming one market.
2. Choose a segment whose actual workflow is accessible.

The matrix retains three mechanisms; this week's experiment investigates reconciliation errors before building across contexts.

Failure case and repair

Turn AI's top ten needs into a roadmap.

The summary conceals source differences and contrary cases.

Trace each judgment to incidents and test the key gap.