Claude Codeループエンジニアリングのすべてのレベル

설명

Learn loop engineering in Claude Code across 3 levels, so AI agents build, verify and ship features while you sleep. Loop engineering explained: the next step past prompt engineering, and the base of real AI automation. Try Hedra at https://hedra.link/ailabs and use code AILABS50 for 50% off your first month. Free setup resources: https://drive.google.com/drive/folders/1pagLScBGfGXkfxYiqCpEkei9KDI8SdCs?usp=sharing Community with All Resources: ailabspro.io/?v=PLyRe6Zk--8 The Roundup, our daily newsletter covering the AI stories that matter. Join now: https://www.theroundup.so/ PASEO: https://paseo.sh/ Open Source Skills in this video: Grill Me: https://github.com/mattpocock/skills GSAP: https://github.com/greensock/gsap-skills Vercel: https://github.com/vercel-labs/agent-skills Supabase: https://github.com/supabase/agent-skills You were already running a loop before you knew what one was. You prompt, the agent builds, you check it, you prompt again. So what is loop engineering? It is handing that checking step to the agent and keeping only the one decision that was always yours: whether the agent is done or needs to keep working. This is the full guide, built on a real salon appointment app that goes from missing a landing page to shipping features overnight. Everything runs in Claude Code, on Warp and VS Code. Before the loop * Why loop engineering did not work a year ago, and what changed with models that run for hours * Why you should never loop your MVP, and what to build by hand instead * CLAUDE.md pointing to AGENTS.md, so the repo is not locked to one agent * design.functional.md for every clickable part of the app * The Grill Me skill, which keeps asking questions until the spec is actually clear Level 1: one loop, one goal (loop engineering in Claude Code) * The naming trap in the slash menu: /loop runs on a timer, /goal runs until the thing is done. Loop engineering is the goal command * How the goal command verifies: a smaller model reads the turn and decides continue or stop * A features folder where each feature holds a spec file and an empty verification folder * The real signal that something needs a loop: it takes a lot of back and forth * A motion heavy landing page with the gsap skill, then the optimize skill to win the speed back * Screenshot based verification pulled from the global CLAUDE.md instead of a full browser every pass * The goal-writer skill, so you never write the "make this a goal" line twice * A loop engineering example: 38 minutes unattended, one error left, and why a screenshot could never catch a blinking mascot Setup: GitHub, Supabase, Vercel * Why your project needs a repo, a database and a deployment, explained for non developers * You make three accounts with Google login and configure nothing * The agent does the rest through each platform's CLI, including the auth step you paste into a second terminal * The Vercel deploy skill and the Supabase skills that auto invoke when the agent needs them Level 2: the software factory loop engineering workflow * Planning several features at once and running them as one queue instead of one feature at a time * Why you build a clickable prototype first: to find out if you wanted the thing you described, and to give the loop something to verify against * The new-feature and functional-ui skills, and the mocks folder that holds a working fake version of the app * The rule that matters most: the agent that does the work never verifies the work * Subagent builds on a branch, adversarial review agent assumes there is a bug, main agent loops until the row is ticked * feature-batch and a queue.md table, with the goal running until nothing is left in todo or building * Pull requests with screenshots attached, plus how to check the branch locally before you merge * Merge to main, Vercel picks it up, and the feature is live for users Level 3: get off the laptop * The only two jobs left: plan the feature, approve the merge. Neither needs a keyboard * Paseo, a free app that runs Claude Code on your own machine and gives you a window into it from your phone, so your skills and logged in CLIs keep working * Why we skipped Claude Code's built in remote control: no skills menu. In Paseo, /goal and every slash command still work * The mobile preview skill, which deploys the HTML mocks as free Vercel links so you can click through a prototype on your phone instead of squinting at images 00:00 Intro 00:46 What A Loop Is 02:08 The Project 04:32 Level 1 07:58 The Goal Writer Skill 08:57 Setup 11:11 Platform Skills 11:53 Sponsor 12:46 Level 2 16:14 The Functional UI Skill 17:51 The Feature Batch Skill 20:13 Level 3 22:34 Mobile Preview Skill Hashtags: #ai #claude #claudeCode #codex #chatgpt #claudeAi #aiAutomation #loopEngineering