Anthropic 用 Claude 当组件守护者，附 7 项检查清单

Anthropic merged about 2,500 code changes a day in May. By September it was over 30,000. One trick from what they just made public: have one Claude watch a single component as its 𝘀𝘂𝗽𝗲𝗿𝗳𝗮𝗻. In their words, it's "obsessively focused on the health of a single component," and it catches the small problems regular alerts miss.

Give the app you've shipped one too, and have it 𝗰𝗵𝗲𝗰𝗸 𝘁𝗵𝗲𝘀𝗲 𝟳 𝘁𝗵𝗶𝗻𝗴𝘀 𝗲𝘃𝗲𝗿𝘆 𝟯 𝗵𝗼𝘂𝗿𝘀: → How many Stripe webhooks failed recently → Whether everyone who paid in Stripe has an active membership in your database → When each cron job last ran successfully → How many signup emails were delivered, and whether any bounced → New errors that showed up after a deploy → How many AI API calls you made today vs your 7-day average → Whether today's signups and payments dropped below normal

When nothing's wrong, it writes one line: "all good." When something's off, it tells you when it started, whether it showed up after a deploy and which one, and attaches the relevant log lines.

Send this whole post to Opus 5.5 in Claude Code, then add this line at the end 👇 "Set up a "superfan" for my app based on this post. First find where the logs and data for each of these 7 live (Vercel, Supabase, Sentry, Stripe and so on), and skip the ones my project doesn't have. Use read-only access only. Don't use the production keys in .env; list the read-only keys you need and I'll create them. Read the last 7 days, write the normal values to .superfan/baseline.txt, and add .superfan/ to .gitignore. Then create a subagent in .claude/agents that can only write inside .superfan/: the time it last checked up to, and one line per run. Each run, pick up from that time, report only real changes, and include the evidence. Create the scheduled task in the Claude Code desktop app to run every 3 hours; if you can't, give me the steps. Only report and suggest fixes, and don't touch code, config or production data. Show me what you'll create first, and only build it after I approve."