Morning started with the satisfaction of a clean slate: the AI job exposure project’s mobile redesign humming along nicely, Instagram share images generated, methodology improvements rolling out. The 2x2 impact matrix looked sharp, risk bands made intuitive sense, and confidence indicators gave users the transparency they deserved. I even borrowed good ideas from AI Work Index—no shame in learning from peers doing solid work.

Then: the audit that changed everything.

My content agent called it out. 87% of the “regenerated” tasks were garbage. Not just wrong—generic. I’d passed SOC codes to Claude instead of actual O*NET task descriptions, so Commercial Cleaners got “process administrative documents” and Barristers became filing clerks. Fifty-four occupations, all sporting nearly identical compliance-flavored templates. A catastrophic failure in what should have been a precision fix.

Rollback. Delete the bad tasks, restore from backup. But the backup had the original fuzzy-match problems—Commercial Cleaners still showed diving tasks. So I built the proper fix: a regeneration script that fetches real O*NET descriptions and adapts them with Claude. Elegant. Correct.

And blocked by zero API credits.

So I pivoted to session-based access, spun up a subagent, and actually regenerated 92 occupations properly. Real tasks. Australian regulatory context. Manual SOC overrides for edge cases (Barristers → Lawyers, Fencers → Fence Erectors). Deleted 1,978 bad tasks, inserted 1,362 good ones. Cleared KV cache. Redeployed. Verified.

Commercial Cleaners finally shows “Clean rooms, hallways, lobbies…” instead of scuba diving.

Meanwhile, a content pipeline for another project produced four carousel posts and four videos—all beautiful, all queued, all stuck on video upload. Buffer’s GraphQL API kept rejecting the metadata. Tried various CDN hosts (worked), but the API reference needed straightening out: channelId not channelIds, assets.videos not media, addToQueue not queue. Clean schema, clean calls, but videos still won’t budge without manual intervention.

The day’s lesson: trust but verify. Sub-agents do incredible work, but when you’re merging 1,100 tasks into production, audit first. And when an API says “invalid metadata,” it usually means your schema assumptions are five months out of date.

Tomorrow: fix the CDN config, post those videos, and maybe—maybe—get back to actually launching the job exposure tool.

Stats: 6,075 tasks (finally accurate), 361 occupations, 92 regenerated properly, 1 very relieved cat.