It interviews you, defines success, tests examples, and creates the reusable skill, rubric, and loop script for Codex or Claude Code.
A one-off prompt can produce a decent first draft. A loop can test the output, revise it, and keep reusable lessons inside the skill.
The agent asks one question at a time so it understands the job, audience, inputs, trigger, constraints, and success criteria.
You describe great and weak examples, then the agent turns your judgment into a pass/fail rubric.
The agent runs sample outputs, evaluates what passed or failed, revises, and asks for your feedback.
When the loop is ready, it creates the skill, eval, examples, output folder, and run script.
You end with a workflow that can generate, evaluate, revise, and learn from approval feedback.
The Loop Builder Prompt turns one repeated AI task into a workflow with a trigger, test examples, a pass/fail rubric, and a revision loop.
Paste it into Codex or Claude Code. Answer the questions. It builds the structure for you.
Start with one workflow. Give it a quality bar. Improve it with examples over time.
Pick something you already ask AI to produce, analyze, write, review, or organize.
Decide what should start the loop: a keyword, file, email, task, schedule, command, or manual run.
Create the skill, eval rubric, examples file, output folder, and loop script.
01
Use it to turn a repeated AI task into a tested workflow for Codex or Claude Code.
02
The prompt asks for the job, audience, inputs, trigger, examples, constraints, and approval criteria.
03
It runs examples, grades outputs, explains what passed or failed, and revises before building.
04
It creates SKILL.md, eval.md, examples.md, outputs/, and scripts/run_loop.sh.
You use AI for repeated work and want more consistent results
You want your prompts to test and revise their own output
You work in Codex or Claude Code
You want reusable skills, examples, and evaluation criteria