Why the context you build up with one AI assistant doesn't move cleanly to the next, and how a local, version-controlled markdown profile keeps you out of any single platform's proprietary memory feature.
A retrospective on using Claude Fable 5 only for first-pass work during its short free window: the tier ladder it sat on, what I built with it, real ccusage numbers against an Opus baseline, and what instrumentation caught that memory would have gotten wrong.
A build log of moving to a Mac, rebuilding the terminal stack, and putting Claude Code, Codex, and Gemini to work on the same codebase. Two real failures included.