When I Started Thinking in Loops: Three Workflows Redesigned as Loops

Do you find yourself re-explaining the same task to AI every single time? You do not have to. You do not need to know how to code. You only need to know how to write rules, and you can get AI to follow your methods systematically. A prompt says "do this for me now" and is forgotten the moment the exchange ends. A rule says "always do it this way from now on," written once, running continuously. This article covers the three loops I built on top of rules, and one first step you can take today.

What a loop actually is

A prompt is one-shot. You ask once, it answers once, and that exchange is over. Next time, you start again.

A loop is different. It is a cycle that starts on its own, runs on its own, and only comes back to you at the moments that matter. You take something you do repeatedly, break it into stages, give each stage a clear input, output, and completion condition, and let it run round after round. The key is that it feeds itself: each round's output becomes the next round's input.

I have written a separate piece on the details of loop engineering (link at the end). For now, just hold onto this one sentence: a prompt waits for you to push it; a loop turns on its own once you have designed it.

How do you get it to turn on its own? The key is writing rules, not code. Describe in plain language how you normally judge and decide, hand those descriptions to AI as rules, and it will follow them round after round. If you can write rules, you can design your own loops.

My three loops

These three loops were not planned from the start. They grew one by one.

The first workflow I turned into a loop was publishing. How does an idea travel from a journal entry and a short post all the way into a full article on my site and then out to other platforms? I broke that journey into stages, wrote it as rules, and let it run. (I wrote a full piece on that loop separately: Using One Article's Workflow to Explain What Loop Engineering Is.)

Once publishing was running smoothly, I noticed two other workflows I was running every day that were equally suited to becoming loops: one for distilling judgments into rules, and one for looking back and improving after the fact. That is how these three loops came to be.

How the three loops fit together

Only after building all three did I notice something: they each point in a different direction. Publishing moves outward, distillation moves inward, retrospection moves backward. Three loops orbiting the same thing: my own knowledge base and system.

Distillation deposits judgments into the knowledge base. Publishing draws from the knowledge base and pushes content outward. Retrospection writes each experience back into the knowledge base. Deposit, push out, write back: with all three running, the knowledge base grows richer with each turn rather than standing still.

Prompts you can copy: let AI help you build a loop

By now you probably understand the concept. But a concept does not become your loop on its own. You need something you can put to use. The three prompt templates below are ready to copy. Paste any of them into your AI agent of choice (ChatGPT, Claude, or Codex all work) and it will get to work.

For each loop, we use the same five dimensions to describe it clearly. Fill these in when designing, and neither you nor AI will miss anything important:

1. Trigger: What event or condition starts this loop.
2. What to read each time: Which logs, examples, rules, or materials to review before each run.
3. How to decide: Which standards govern choices when the loop reaches a decision point.
4. What counts as done: The verifiable state that marks the end of a run.
5. Human checkpoints: Which steps require your sign-off before AI continues.

I want to turn the repetitive work I do into loops that can run repeatedly. Please handle two scenarios:

A. If you can see my work logs, past records, or examples, start by reviewing those and identify three tasks I have been doing repeatedly with roughly consistent steps.
B. If you cannot see any records, ask me to describe three tasks I do repeatedly, and for each one ask: what is the goal, what information is needed, and how do I normally judge whether it went well.

Then use these five columns to organize the three tasks: Trigger, What to read each time, How to decide, What counts as done, Human checkpoints.
Under "How to decide," you may infer from my past patterns, but mark each inference as "pending my confirmation."
Finally, recommend the one loop most worth building first and explain your reasoning.

Please help me design "(fill in the workflow you want to build)" as a loop.
Output using these five columns: Trigger, What to read each time, How to decide, What counts as done, Human checkpoints.
Write the rules in my own voice. Ask me any necessary questions before starting, then produce a version I can copy and use directly.

Please review the recent outputs from this loop, along with my corrections and comments on each run. Identify the places where AI made consistent errors, missed things, or produced text I rewrote.
Organize those findings into new rules, each written as "whenever X, do Y."
Then list which rules need my confirmation before being added to the loop officially.

Not every loop needs to run all the way through

One clarification before closing. Designing something as a loop does not mean you always need to run the full cycle. A loop is a map showing you where you could go, not a conveyor belt that cannot be stopped.

If I only want to capture one thought, I stop at the journal. If I want to share something briefly, I stop at the short post. Only when something is genuinely worth developing further do I carry it up to a full site article. Stop when you should. Pull back when returns are diminishing. Do not run a loop to completion just for the sake of completing it.

The practical benefit a loop gives me is simple: I do not have to start over every time I do the same thing. For knowledge workers, that translates to moving time away from repetitive effort and toward work that actually accumulates. Tools, platforms, and models will keep changing, but the ability to take what you do repeatedly and design it into a loop that runs on its own and feeds itself back: that capability will always be worth having.

If you want to start, do one thing today: write down the instruction you most often find yourself re-explaining to AI. That is your first rule. You do not need to know how to code. If you can write rules, you are already designing your own loops.