A Plain Memo
What I learned from 2,279 commits of AI mistakes
If you want a system that runs itself, you don't design the safety net — you pay for it one failure at a time, and the trick is writing down the receipt so the same mistake can never bill you twice.
SPECIAL EDITION — the making of Plain Memo. This is the internal archive of how the channel itself was built. Nine weeks, 2,279 commits, hundreds of documented failures, and the self-healing machinery that emerged from every scar — all logged, priced, and turned into permanent guardrails. This episode is produced BY the automated pipeline it describes.
ℹ️ AI voice disclosure: This episode uses an AI-generated voice (ElevenLabs). The host is a pseudonymous byline persona (Reid Conlin).
In this memo
Sources
- Nine weeks of production engineering logs with costs dated and measured (2026-05-12 to 2026-07-16)
- Shipped concerns archive: 622 documented failures with root causes and fixes
- Model pricing analysis: 184 AI video endpoints evaluated for cost and quality
- Real YouTube audience retention data from early episodes (35-45% viewer drop in first 30 seconds)
- The automatic clean-room recovery system that emerged from Episode 18's five phantom failures
- Git history: 2,279 commits across the nine-week build