/goal & /loop as Verifiable AI Loops
Loop Engineering is not a magic command. Learn how to design verifiable, stoppable AI loops with /goal, /loop, harness, verifier, and stop rules for reliable automation.
“AI Disruption” Publication 10,200 Subscriptions 20% Discount Offer Link.
Loop Engineering has been discussed by many people recently, but it is not a new magic word, nor is it as simple as prefixing a prompt with /loop to make the AI automatically deliver results.
I prefer to define it as: designing a repeatable, verifiable, stoppable, and resumable work loop for AI Agents.
This article will clarify several key concepts:
What tasks
/goaland/loopare each suited forWhat Harness, Verifier, Memory, and Stop Rules actually do inside a loop
The corresponding capabilities of Claude Code and Codex in these areas
How to write a practical, implementable loop task specification
Which pitfalls turn loops into infinite loops, fake completions, or cost black holes
Why We Should Take Loop Engineering Seriously Now
In the past, when we used AI for programming, most interactions were single-turn:
I write a prompt.
AI modifies the code.
I run tests, check errors, and paste them back.
AI fixes them again.
Repeat.
This process may look like the AI is working, but in reality, the human is constantly acting as a manual scheduler inside the loop: deciding the next step, checking output, determining whether to continue, and recording what has been changed.
What Loop Engineering aims to solve is not “making the AI give prettier answers,” but explicitly designing these actions:
What the goal is
What one round should do
How to verify each round
When to stop after failures
How to determine success
Where to store intermediate state
Which permissions can be executed automatically and which require human confirmation
This is also the fundamental difference between Loop Engineering and ordinary Prompt Engineering.
Prompt Engineering focuses on “how to make the model give a better response this time.”
Loop Engineering focuses on “how to make the model continuously move toward a verifiable result across multiple rounds of execution.”
If you’re only asking for a single concept explanation, a regular prompt is sufficient. But if you want the AI to continuously modify a module, write tests, fix builds, research information, update documentation, and repeatedly verify — that’s when Loop Engineering starts to show its value.
