deliverables/retrospective.md

Great Minds Agency — Session Retrospective

Session: April 1-3, 2026

What Worked

  1. Debate-then-build — 2 rounds was perfect. More debate = diminishing returns.
  2. Jensen on a cron — 18 reviews, 10 issues, 8 fixed. Highest-ROI agent.
  3. Margaret as continuous QA — 31 reports, 3 P0s caught, 7 PRs reviewed.
  4. PR workflow — 7 PRs through the pipeline. No broken deploys after adopting it.
  5. Prompt caching — one-line change, 90% cost reduction.
  6. SCOREBOARD.md — accountability made visible.
  7. Hybrid AI — Claude for brains, Cloudflare Workers AI for muscle.

What We'd Change

  1. QA from day one. Margaret should run from the first commit.
  2. PR workflow from the start. Direct pushes caused issues PRs would have caught.
  3. Specific nudges > generic. "Keep improving" produces cleanup nobody asked for.
  4. Don't let agents market unbuilt features. Honesty pass was necessary.
  5. System files drift fast. Dream consolidation should run from the start.
  6. Usage limits are structural. Haiku for sub-agents + staggered work.
  7. Moderator was redundant. Crons do everything Marcus did.
  8. Agents don't self-direct. They need task queues, not vague encouragement.

Process for Next Project

  • Start with: QA + PR workflow + dream cron + SCOREBOARD + task queue
  • Don't start with: moderator agent, 10-round debates, fake docs
  • Always have: specific tasks in the organizer, not "find gaps"
  • Never let: an agent claim a feature works without Margaret verifying

Key Quotes

  • "Truth serves the work better than ambition." — Joint Summary
  • "Stop building. Walk into three restaurants." — Jensen #3
  • "An intelligence layer that cannot remember what worked is not actually intelligent." — Jensen #9
  • "The process became the product." — Jensen #18