6 LOOPS IN LIBRARY · GROWN BY ITS OWN LOOPSNETWORK:open-prompts3dtextgenerator
Loop engineering · est. June 2026

Stop prompting
your agents.
Design the loops
that prompt them.

Copy-paste loop templates for AI coding agents — Claude Code, Codex and OpenClaw — each with a goal, a cron trigger, a verify step, an exit condition and a real token cost.

7 fields per loop3 agent runtimes100% open source
content-refresh.looprunning
DISCOVERACTVERIFYPERSISTDECIDECYCLE#0247
TRIGGER 0 3 * * * · EXIT rank_delta ≥ 0~41k tok/cycle

What is loop engineering?

Loop engineering is the practice of designing systems that run your coding agents on a schedule — discover work, act, verify, and persist state without you in the loop. These agent loops replace one-off prompts with repeatable cycles that have exit conditions, verify gates, and measured token costs per cycle. Whether you run Claude Code, Codex, or OpenClaw, agent loops are how production teams ship autonomous workflows instead of babysitting every prompt. Every template on this site follows the same seven-field spec so you can audit any loop in under a minute. Read our full explainer on loop engineering →

The loop format

The loop engineering format — seven fields, one standard

Every loop on this site is written in the same seven-field format, so you can audit any loop in under a minute.

F1

Goal

One sentence. What "done" means, stated so a model can judge it.

F2

Trigger

Cron expression or event. When the loop wakes up.

F3

Discover

Where tasks come from — a queue, a diff, an API, a board.

F4

Act

The agent call. Model, tools, and the prompt the loop writes for it.

F5

Verify

The automated check that gates every cycle. Tests, lint, rank data.

F6

Persist

State that survives the run — markdown file, board, or table.

F7

Exit

The condition that stops the loop. No exit, no loop — just a bill.

→ Every template also declares est. tokens per cycle, because the loop is the new cost center.

Running a loop we haven't catalogued?

Submit it in the seven-field format. We test-run every loop before it's listed, and credit you on the template page.

Submit a loop →