Claude Code tutorial: build and improve coding loops

AI Labs explains a Claude Code workflow that builds one feature at a time, checks the result, and records each round. The tutorial adapts Karpathy's experiment loop: an agent changes code while a separate scoring mechanism stays outside its editing permissions.

The setup uses a project context skill as an app memory bank. A build skill coordinates the work, and a write checks skill defines tests before implementation. After human review, an approval program moves the checks into a folder protected by Claude Code settings. The build skill then commits the approved checks. A fresh feature builder agent handles each feature, with instructions in program.md and results in a separate file.

The examples show why passing checks can leave work unfinished. AI Labs reports that a restaurant ordering feature passed 11 checks before the loop added and connected the customer order form. In a project management app, a shared database passed 10 checks even though the app saved nothing to it.

A second skill, auto loop, reviews results and rewrites the working instructions for later features. It cannot edit the checks. The speaker says this caught recurring integration mistakes and conflicts with existing mentions behavior. For anyone configuring a coding assistant, the practical constraints are repeated tasks, sufficient token allowance, automated checks, and an environment where the agent can run its work.