@hackproduct9: Most people learn Claude Code's agentic loops backwards. They learn how to start one. Nobody teaches them how to stop one. Same failing test in all four. Watch what changes. 🔁 Turn-based — closes on you. Claude reads, edits, tests, hands back, waits. You review the diff and write the next prompt. fix the failing test in cart.test.js 🎯 Goal-based — closes on a judge. After every turn a separate fast model (Haiku) reads your condition plus the transcript and returns yes/no. No? Back to work. /goal all tests pass, stop after 5 tries ⏱️ Time-based — closes on the clock. Nothing is proven. A timer wakes Claude, it works, it sleeps, it wakes again. /loop runs on your machine. /schedule moves it to the cloud — laptop can be shut. /loop 10m run the tests, fix what's red 🌙 Proactive — nothing closes it. It just runs. An event fires it while nobody's watching. Claude spawns subagents to triage, fix and review, then re-arms for the next event. /schedule 1h: red CI, fix it, open a PR 🔴 "red CI" — CI is Continuous Integration, the robot that runs your test suite on every push. It reports a colour. Green = passed. Red = broke. So red CI just means the build is broken. That's what wakes the agent at 3am. 👀 The part nobody separates: /loop and /goal look identical and are opposites. /loop starts the next turn when a timer elapses. /goal starts it when a check hasn't passed. Clock vs evidence. Use /loop to watch something, /goal to finish something. The rule: pick your loop by its exit, not its trigger. Most people reach for /loop when they meant /goal, then poll every 5 min for a refactor that finished in 90 seconds. Send this to whoever has one running right now. 📸 Screenshot the last frame — four loops, four commands, four exits. Follow @hackproduct — scary AI concepts, made shippable. ⚡ . . #claudecode #AIagents #agenticAI #AIengineering #softwareengineering loops

HackProduct

Open In TikTok:

Region: US

Wednesday 05 August 2026 22:36:57 GMT