Double Loop

Part of the cognition series.

Every magic trick has three acts. The pledge, where you show something ordinary. The turn, where you make it extraordinary. The prestige, where you bring it back — changed.

The Pledge

You are using AI as a pipe.

Prompt in, text out, done. You read the output, maybe fix a typo, and ship it. The writing is fluent and says nothing that couldn’t be said by anyone else with the same prompt.

This is how most people use AI. It’s also how you use a vending machine. Insert coin, receive product. The machine doesn’t know what you wanted. It knows what button you pressed. The gap between those two things is the work, and you skipped it.

The output reads well on first pass. That’s the trap. “Good enough” has a specific texture: em dashes in place of ideas, the same point restated three ways, stock metaphors paving roads to nowhere. You used the pipe once when you needed to use it six times.

The Turn

In 1967, Gregory Rabassa received the manuscript of One Hundred Years of Solitude from Gabriel García Márquez. The Spanish was lyrical, compound sentences running for paragraphs, rhythm doing half the work of meaning.

Rabassa read a sentence, heard the rhythm in Spanish, felt for the English rhythm that carried the same weight, tried a rendering, read it aloud, killed it, tried another. He kept the verb from one attempt and the cadence from another. He revised until the English sentence didn’t say what the Spanish sentence said — it did what the Spanish sentence did. Then he moved to the next sentence.

García Márquez reportedly said the English was better than his original.

That was not a pipe. That was a loop: read, parse, try, reject, try again, keep. Six steps. On every sentence. Rabassa ran all six inside one mind. The same structure holds when the labor splits across two minds with different strengths.

Here are the same six operations any editor already performs, whether they name them or not: perceive, cache, filter, attend, consolidate, remember. Either party can perform each one. One usually leads.

Perceive. You see what nobody else sees. The connection spanning three disciplines, the lived experience that no training corpus encodes. You bring the raw insight.

Cache. The AI organizes it. A million tokens of context, perfect recall, no ego about the outline. It will find an order you wouldn’t have.

Filter. You kill what doesn’t belong. The AI will give you five paragraphs where two would do. Delete the three. The AI can’t do this. It doesn’t know which line is yours and which is filler.

Attend. The AI surfaces what you missed. It reads every word with equal attention, no attachment to the clever phrase you spent ten minutes on. Let it question your filter. Have the courage to hear it.

Consolidate. You tighten. Each pass compresses. The AI can compress too — say “make this half as long” and it will — but you decide what it lost. Sometimes the thing it cut was the point, and now you know what the point was.

Remember. The AI preserves what the loop learned — the decisions, the failed attempts, the editorial memory that makes the next pass smarter. A draft in your notes is a thought you had. A published piece is a thought that exists in the world.

Then run it again. The published piece becomes the next cycle’s input. You perceive it fresh, notice what’s weak, and the loop continues.

The attitude is hansei: not the humility that yields, the humility that reciprocates. You will never have a million-token context window. The agent will never see the connections spanning across disciplines.

The people who say AI makes them dumber are automating perception and publication while skipping judgment. The thinking was never in the typing.

The Prestige

The method is learnable. The skills are open. The models are available to everyone. What none of that replaces is courage.

Chris Argyris called it double-loop learning: the first loop corrects errors, the second loop questions the assumptions that caused them. Single-loop editing fixes the sentence. Double-loop editing asks whether your taste in sentences is any good.

Taste matters. But taste is courage at a tighter loop. Saying “this line is good” is an identity declaration, a vulnerability moment. Choosing not to judge is the cowardice that keeps people in the pipe. To develop taste, you have to admit your current taste might be wrong, then run the experiments. That cost is multiplicative — courage at the outer loop times courage at every inner loop that tunes the judgment.

While writing The Natural Framework, I built a six-step pipeline across twenty-five domains. Then I asked Claude: will any organism that breaks any step in the loop, die? I did not know the answer. If prions didn’t compound, or if p53 loss didn’t cascade, it would’ve falsified the framework that I agonized over all morning. They died. Five times. The courage was in asking before I knew.

Rabassa had the same thing. He killed his own renderings until the English did what the Spanish did. The loop teaches you to write by forcing you to edit. But the loop can only start if you’re willing to be wrong.

Agents don’t have agency. Humans do. Outsource judgment, and the agency goes with it.

Write something that only you could write. Run it through the six steps until it stops getting better. Then publish it where other minds can perceive what you remembered.

PageLeft was written for that.

Written with Claude Opus 4.6 and GPT-5.4 via Codex. The process is the post.