← home

Why Your AI Needs a Loop (And a Skill)

2026-05-22

You already know how to iterate with AI.

You type a prompt. It gives you something close. You adjust. It gets better. You adjust again. Eventually it's right.

That back-and-forth isn't new. You've been doing it since ChatGPT came out.

But here's what most people miss: that back-and-forth can happen without you.

The ChatGPT Dance

This is what everyone does today. You and the AI volley until it's right:

 You:  "How much revenue did we do last quarter?"

 AI:   "Here's a breakdown by month..." 
       ❌ (wrong — you needed by region)

 You:  "No, I need it broken down by region."

 AI:   "Here's Q3 revenue by region..."
       ❌ (wrong — you said Q4)

 You:  "That's Q3. I said Q4."

 AI:   "Here's Q4 revenue by region..."
       ✅ (finally right, after 3 tries)
      

Three prompts from you. Three responses from the AI. That's the standard experience.

The user is driving the iteration. The user is the one saying "no, that's wrong, try again."

Now Flip It

What if instead of you doing that back-and-forth, the AI did it with itself?

Same task. Same model. Same tools. But this time, you send one prompt and walk away.

Inside the system, the AI generates a query, checks the result, realizes it's wrong, tries a different query, checks again, combines findings, validates, and only when it's satisfied does it return one single answer to you.

 You:  "How much revenue did we do last quarter?"

       ────────────────────────────────────────────────────
       AI goes to work — you do nothing
       ────────────────────────────────────────────────────

  Query 1: by month → "Does this match the ask? No."
  Query 2: Q3 by region → "Does this match the ask? No."
  Query 3: Q4 by product → "Does this match the ask? No."
  Query 4: Q4 by region → "Does this match the ask? Yes. ✅"

       ────────────────────────────────────────────────────

 AI:   "Here's Q4 revenue by region..."
       ✅ Right on the first response to YOU
      

Fundamentally, it's the same thing. There's still iteration. There's still trial and error. There's still correction.

But the AI is doing the prompting, not you. It's going back and forth with itself instead of with you.

One prompt in. One answer out. Zero back-and-forth with a human.

How the Loop Works

All agents have a finish reason. It's a signal that says why the agent stopped: did it call a tool? Did it finish answering? Did it hit a limit?

I used that finish reason to build a loop. The agent calls a tool, gets a result, checks if the result satisfies the original request, and uses the finish reason to decide whether to keep going or stop.

But Even Loops Fumble

Here's the catch: even with a loop, the AI still has to figure out the right approach every single time.

Ask it a SQL question today, it tries four queries before getting it right. Ask a similar question tomorrow, it tries four queries again.

That's not intelligence. That's persistence.

From Tools to Skills

This is where the real leap happens.

You can hand a carpenter some two-by-fours, a hammer, nails, a nail gun, and an air compressor. Those are tools.

But until you teach them how to frame a wall, they're just holding equipment. Once they know the sequence, the spacing, the technique — it's now a skill they carry with them.

Or think about IKEA furniture. You get pieces, fasteners, and tools. But you also get instructions. The instructions — the script — that's the skill.

For an AI, tools are the functions it can call. Skills are the learned procedures that tell it which tools to use, in what order, and with what parameters — because it's already done the trial-and-error once.

The first time, it might try four SQL queries before getting it right. The second time, with the skill captured, it knows which query pattern works for that type of question.

 FIRST TIME (loop only, no skill)

  You:  "Q3 revenue by region?"
       ↓
  Query 1 → wrong columns → retry
  Query 2 → wrong filter → retry
  Query 3 → wrong join → retry
  Query 4 → correct ✅
       ↓
  4 iterations. You waited. It got there.


 SECOND TIME (skill applied)

  You:  "Q4 revenue by region?"
       ↓
  SKILL: revenue_query_pattern
       ↓
  Query 1 → correct ✅
       ↓
  1 iteration. Instant.
      

The skill is the instruction manual built from real experience. Not generic documentation — the specific sequence that worked.

What This Changes

Most people think the next breakthrough in AI is a bigger model with more parameters.

It's not.

The breakthrough is architecture. One prompt in. The AI iterates with itself. One answer out. Then capture that iteration as a skill so it never has to relearn it.

And I think it can go further: one-in, many-out. Many-in, one-out. Zero-in, many-out.

The model doesn't change. The way you use it does.

— Jon