AI can now generate polished work from remarkably little direction.
That is impressive. It is also dangerous.
A polished response can create the illusion of finished thinking. You can accept it and move on without understanding the reasoning beneath it.
That is not leverage. It is abdication.
Effective AI utilization has moved beyond clever prompt engineering toward context engineering. Anthropic calls context engineering the natural progression of prompt engineering. Prompt engineering focuses on effective instructions. Context engineering is broader: curating the information, sources, constraints, examples, and goals the model needs throughout the work.
This matters because the quality of the middle depends heavily on the thoughtfulness of the beginning.
Why I Rejected 10-80-10
I previously encountered a 10-80-10 framework for working with AI:
The first 10% is human direction.
The middle 80% is AI execution.
The final 10% is human review.
The structure made sense. But I challenged the allocation.
Ten percent is not enough for the consequential work that context engineering often requires. It also understates the human judgment, verification, and iteration needed before something is ready to ship.
My alternative is 20-60-20.
This is not a scientific formula for every assignment. It is a practical operating model that keeps humans responsible at both ends while AI creates leverage in the middle.
The First 20%: Engineer the Context
Before asking AI to produce, define the problem, audience, desired outcome, source material, constraints, risks, and what excellence looks like at the finish line.
Context engineering is not simply writing a longer prompt. It requires thinking clearly about the assignment before delegating any of it.
If you cannot define the finish line, AI may lead you down the wrong path.
The Middle 60%: Create Leverage
Once the direction is clear, let AI carry much of the production burden.
Depending on the assignment, it can research, organize, generate alternatives, analyze arguments, build drafts, identify gaps, and compress material.
This is not necessarily a single prompt followed by a single answer. The middle can include multiple loops in which AI produces, evaluates, and improves its work while you continue steering.
The Final 20%: Protect Understanding
Andrej Karpathy posted a warning that belongs in every leader’s AI operating system:
“You can outsource your thinking, but you cannot outsource your understanding.”
A polished result is not the same as genuine understanding.
The final 20% is where you challenge the logic, verify claims, resolve contradictions, add experience, refine the message, and ensure you can defend the work without AI in the room.
A 2025 Microsoft Research study of 319 knowledge workers found that greater confidence in generative AI was associated with less critical thinking. It does not prove AI causes long-term cognitive decline, but it reinforces the risk of accepting AI output without active engagement.
Never ship work you cannot explain or defend.
What 20-60-20 Looked Like in Practice
I recently needed to create content for a group of leaders. Based on similar work, I estimate it would normally have taken 90 to 120 minutes.
Using AI and the 20-60-20 rule, I completed it during a single 30-minute focus block.
I wrote or spoke approximately 2,900 words of context and direction across 18 iterations. AI helped organize, challenge, draft, and compress the material. During key iterations, I used Read Aloud to hear the content as someone encountering it for the first time would. The finished piece was 514 words.
When I delivered it, no one had any questions, comments, or suggested improvements. It accomplished the finish line I had established at the beginning.
AI did not eliminate my involvement. It concentrated my involvement where it created the most value.
Learning That Becomes Leverage
Through my Leadership Recall Learning Operating System, I had captured both 10-80-10 and Karpathy’s warning because I recognized they could prove useful later. When the ideas intersected, I could recall them, challenge the original framework, and build a better way of working with AI.
That is the larger discipline: continually learning, building, experimenting, and refining.
Consuming more information is not the advantage. Turning retained knowledge into better judgment, stronger systems, and improved performance is.
Test It Yourself
Use your next meaningful assignment as a 20-60-20 experiment:
Engineer the context and define the finish line.
Let AI handle the core execution using structured loops.
Challenge, verify, refine, and ensure you understand the result.
For more on using loops in the middle 60%, see Five Building Blocks to Close the AI Gap Faster.
Then ask:
Do I understand this well enough to defend it without AI in the room?
Give AI the middle. Never give it your mind.
If you want to build practical AI workflows that increase leverage without surrendering judgment, join us inside Special Ops in the Elite Ops Academy. Reply if you're curious to know more.
Own the outcome,
Steve Kahle | ELITE OPERATORS
In Pursuit of Elite
