They start by insisting on a six-step audit of every AI-agent session stored on the machine, from Claude and Codex logs through any JSONL in config or local-share folders. First, you list each source—tool name, path, file count, size, date range—and propose which to sample. Only when that inventory and a sampling plan are approved do you move on. Next comes distillation: you never load entire logs, only 150 lines from up to 200 files, chosen as the 15 newest, 10 oldest, and 20 evenly spaced. You grep for corrections (“no, I meant…,” repeated requests, tantrums) and dump findings into an evidence.md broken into six sections—Recurring themes, Abandonment graveyard, Correction patterns, Repetition tax, Rhythm, Blind spots—with every claim backed by at least three dated quotes. Phase 3 tests five hidden hypotheses by asking one revealing question each, comparing answers to the evidence and drilling into any contradictions. In Phase 4, the Mirror, you present raw patterns: what they truly believe (hours spent), where their mind loops, strengths, avoidances, and one unheard truth—each linked to receipts, hardest point last, stated once. Phases 5 and 6 turn insight into action. Leverage lists where time dies, what to delegate to AI (ranked by hours saved), what only they should do, what to drop, what to keep, and when to schedule tasks. Residue produces three artifacts—mirror.md, roadmap.md, and proposed config diffs for agents—each reviewed and approved before changing any files. The prompt demands plain speech, precise receipts, no speculation, and keeps all data on the local machine.