shipped spoolcast dev-log 7 — the AdsMetri three-agent build

shipped the AdsMetri devlog about building one product across Codex, Manus, and Claude.

the useful lesson was not "use three AIs." that version is too clean and mostly false. the product only started making sense when the roles separated: Claude for design/spec direction, Manus for foundation and live verification, Codex for repo-level implementation and debugging.

the video also exposed a few pipeline problems while making it. thumbnails needed deterministic text overlays instead of asking the image model to spell. mobile export needed clearer aspect defaults. audits were useful, but too many repeated audit passes turned into expensive procrastination with better logging. not catastrophic. just the usual reminder that automation still needs rules, otherwise it becomes a very fast way to make the same mistake five times.