Project 01 of 07 · Imagera AI · Time back
Eight Agents, One Review Gate
Drafting 45 → 8 minutes · first-pass approval 55% → 88% · creative output 4×.
The numbers
Before, and after. Every figure is on her résumé.
458Minutes per draft55%88%First-pass approval1×4×Creative output
The case
What it cost before. And the system that replaced it.
The cost of waiting
A draft took 45 minutes and only 55% passed first review. Output was capped by hand work.
- Goal
- Launch-grade marketing output every day, without adding headcount or letting quality drift.
- What I built
- Eight marketing agents in Claude Code: Agent Skills hold the brand rules, MCP connects Meta Ads, GA4 and Ahrefs, a golden set scored by an LLM judge checks every change, and a person approves before anything ships. Each agent has an owner, an eval set and a rollback path; a weekly review keeps first-pass quality above 85%.
- 01Brief
- 02Brand skill
- 03Agent draft
- 04Golden-set eval
- 05Human gate
- 06Publish
- Claude Code
- Agent Skills
- MCP
- LLM-as-judge
- GitHub
Why it works
The default. Pre-select the plan most customers should choose.
The default decides most outcomes. Here the default is safety: nothing ships until a person says yes, so the team trusts the agents enough to let them do the volume.
Where
Imagera AI
AI Marketing Lead · Contract, full-time · Jan 2025 – now · Canada · remote
A growth-stage AI product. Led a pod of 5.
