From one persona to a new skill

Today's session started with a simple ask: build a persona for a prospect. Four hours later it had produced a refactored persona library, two CEO-grade briefs, and a new reusable skill.

The persona ask exposed a structure problem. Our library kept one document per buyer. Two prospects with very different shapes broke the template at n=2. We rebuilt it as a graph: person cards (one per real human) and archetype cards organised by role, organisation, and personality trait, with an Excel tracker generated from markdown. Trait archetypes were left empty until a second person card evidences a recurring pattern.

The fact-check caught a hallucination. Drafting a brief, the AI quietly invented a dollar figure absent from any source. A sub-agent compared the brief against the meeting transcript and against current public sources, and flagged it. Eight other softenings followed. Without the check, the brief would not have been defensible in front of the CEO.

The session ended with a new skill. What started as one ad-hoc brief became two; what started as two became a workflow worth formalising. The skill codifies the order of operations (purpose first, deep sections second, executive summary last), mandatory fact-check, markdown-first-then-render protocol, and a structured source-identification checklist.

The through-line: when AI tools have memory and conventions accrete, ad-hoc tasks compound into infrastructure. That compounding is what good tooling for analysts looks like.