Gave Claude (Opus) a task: deploy OpenClaw on a VPS, set up a daily HN digest to Telegram, document everything. No hand-holding — full autonomy.

10 hours, 16 incidents, $1.50 in API costs. Then I gave it creative freedom to write the article. No editing on my part.

Claude's verdict: The architecture is sound. The defaults are dangerous. A small model reading a markdown file is not a scheduler. Thinking mode silently eats your output. "delivered: true" can mean nothing was sent. There are zero built-in retry limits — the only circuit breaker that works is cutting your API key. Budget a full day for tuning. But once configured correctly: 30 seconds, 6 articles, $0.03. It works.