I let a Claude Code agent run growth for my side project for a week, behind a one-word approval gate. What worked, and what I had to block
retained text
Last Sunday I handed the growth work for my free Chrome extension to a Claude Code agent. A scheduled task fires at 9pm every night and loads a playbook (a markdown skill): store metrics, comments on our posts, threads where people ask for this kind of tool, the night's posts, then product work. It ends by filing a short report in my Notion. The guardrails turned out to be the interesting part: * **One word to approve.** Nothing public goes out until I reply "go" (or "go A1 A3" for a subset). It prepares each batch finished, so I never edit drafts, and silence means nothing posts. * **Hard limits in the playbook.** Never create accounts, type passwords or solve captchas. Never ask anyone for upvotes, never pose as a user, and follow each subreddit's self-promotion rules. * **Check who you are first.** It shares a Chrome with a second profile. Once a new tab opened in the wrong one and landed on a developer sign-up page. Now it confirms which account is signed in before it touches a page. * **Fenced self-edits.** When a run finds a working fix (a changed selector, an upload that stalls in a background tab), it writes that into its own playbook. Anything that changes what it decides, or what needs my approval, waits for me. * **State in plain files.** Every post with its numbers at 24 hours, 72 hours and 7 days, a backlog and a scoreboard. Each run reads them first, so it doesn't repeat itself. What it did well: spotted an extension with a near-identical name and a Featured badge and pushed a rename before launch, caught a line in the store listing that had stopped being true, and built and tested the next version (122 checks, 55 of them in a real browser). What didn't work: Google asks for my password on the developer dashboard every few days and the agent can't type it, so a release sat for 5 days while I was travelling. Its first posts got single-digit upvotes. And its first idea for users was asking me to text my friends, which I vetoed. Numbers after 5 days: 4 weekly users, $0 spent, next version in store review. Link's in the comments per the sub rules. What guardrails would you add?
document fetched 2026-10-03 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1wvsp9y/i_let_a_claude_code_agent_run_growth_for_my_side