Anthropic's Loop Engineering Guide for Claude Code

Summary

Anthropic's loop-engineering guide reframes Claude Code usage as designing loops that run until a stop condition is met, rather than prompting turn by turn. It defines four loop types by autonomy level — turn-based, goal-based (/goal), time-based (/loop, /schedule), and proactive (schedule + goal + skills + workflows) — and stresses that the turn-based upgrade is to encode verification steps in a SKILL.md so the agent checks its own work.

Key Details

• Four loop types map to an autonomy ladder; most daily work still wants a single well-scoped prompt.
• The load-bearing pattern: bake measurable self-verification into SKILL.md so the agent stops being the verification step (guide's example is a browser-verify loop for UI changes).
• Loops cost more — a /goal run used ~3× the tokens of a single turn; pilot on a slice, set turn caps, route routine loops to cheaper models.

Why Rolf Thinks This Matters

Further Reading

• Anthropic Loop Engineering Guide Just Changed How I Use Claude Code