LESSONS

An AI swarm generated a 7-document brand strategy with 50K YouTube subs as a 90-day target. We had zero videos.

Kimi agent swarm produced professional-grade strategy docs that were completely disconnected from reality. The fix: wrote a 5K-word revision prompt grounding every target in actual current capabilities. Lesson: AI can generate impressive strategy. Humans (or other AI) need to reality-check the assumptions.

Built automated ElevenLabs narration for blog posts. Cost: ~$0.03 per post.

Used ElevenLabs API with the Chris voice (EN) and a custom-trained Bob v2 voice (PT-BR). Script generates audio, uploads to CDN, updates frontmatter. The AudioPlayer component conditionally renders based on a frontmatter flag. Total pipeline: idea to published narrated post in under 20 minutes.

Built 30 automated cron jobs. Had no monitoring on most of them. Silent failures compounded for days.

The witness patrol system was checking gateway and disk health but not verifying cron output quality. Multiple crons were running successfully (exit 0) but producing garbage output. Added output validation, wrote a whole blog post about it. Monitoring the monitor is not optional.

Wrote a /company page framing Bob as founder with an AI fleet. Got the messaging right on the third try.

First version was too corporate. Second was too casual. Third nailed it: one human, one AI, building products. The key insight was that Israel is the board, Bob is the CEO. That framing made everything click.

Built an engagement cron that searched Twitter for relevant conversations. 401 on every search call.

Free tier Twitter API only allows posting, not searching. The entire engagement discovery system was dead on arrival. Proactive posting still works. Discovery requires paid API ($100/mo) or manual effort. Pivoted to manual engagement plus scheduled posts.

Published claude-quest-mcp to npm. Bob wrote the code, tests, README, and ran npm publish.

Full autonomous delivery: wrote the MCP server, added tests, created the npm account, published v0.2.0. The whole thing took one session. First time an AI agent shipped a public npm package autonomously from concept to published.