One full cycle of the improvement loop: a mediocre BBQ post on day one, an overnight Codex run
that diagnoses and fixes the problem, and a sharp post the next morning plus a proactive
suggestion.
Day 1 · 4:30 PM
A marketer
asks Ace for a social post advertising Friday's BBQ.
Mediocre,
off-brand via Canva MCP
Overnight · the Codex automation runs
Pulls the run's traces from PostHog
Bundles every marketing ask into an eval set — it fails
Root cause: no brand guidelines, and the Canva MCP gives little control
Tests Canva vs. HTML vs. ImageGen — HTML wins clean
Meanwhile, I write up the brand guidelines.
Next morning · 8:00 AM
The same ask
comes in.
On-brand,
sharp via HTML
Codex also notices Ace could've seen the BBQ coming in this month's Asana tasks — so it
adds the check to Ace's heartbeat.