Contents Jump to section
Beyond the perfect prompt
The machine is learning our language. We still have to explain our world.
Much of the advice around using AI begins with the prompt. Use this structure. Assign this role. Include these instructions. Find the wording that makes the model produce a better answer.
Some of that advice is useful. But it can also make working with AI feel like learning another machine language, just one disguised as English.
Naval Ravikant offers a different approach.
In A Motorcycle for the Mind, he describes speaking to AI in structured English rather than chasing short-lived techniques. His reasoning is that the technology is adapting to people faster than people can keep adapting to its interfaces.[1]
His advice is to “let the AI learn how to be useful to you.”[1]
That does not mean understanding the work no longer matters. In the same conversation, Naval explains that his knowledge of computers and programming helps him specify what he wants and recognise where the technology falls short.[1]
We may need fewer specialised commands. We still need to know what a good result looks like.
In A Return to Code, he describes how forgiving communication with an agent can be: “You can use different words; you can make spelling mistakes; you can explain things your own way.”[2]
For a professional who knows their work but does not speak in technical specifications, that is a meaningful opening.
What the words leave out
An agent can understand a request without knowing enough to do the work well.
“Write a proposal” is clear enough to produce a document. It does not explain the relationship with the client, which promises the business can keep or what went wrong in the previous engagement.
Those things may be obvious to the person asking. They are not automatically available to the system.
Naval touches on this difficulty when discussing his own software projects. “You have to know what you want,” he says, describing that as the hardest thing.[2]
AGL’s reading goes beyond describing the desired result. A useful agent also needs to understand the reasons behind it.
Why does this client need a careful explanation rather than a stronger sales pitch? Why is the company choosing slower growth? Why should a particular conversation remain with a person even if the agent could handle it?
The answers reveal how someone judges their work.
This is the role of Creative Context: making a person’s purpose, values, experience, relationships and way of working clear enough to guide an agent.
A founder’s insistence on plain language may come from years of watching clients misunderstand technical promises. A preference for slower growth may reflect a commitment to protect service quality. Keeping a conversation human may be about trust, not the difficulty of the task.
Without those reasons, the agent may produce something efficient that misses what matters.
The point is to help it understand why one otherwise reasonable answer is right for this situation and another is not.
Learning through the conversation
You do not have to know how to explain everything before you begin.
Often, you discover what you mean by seeing what the agent gets wrong.
A draft can be accurate and still miss the point. Explaining why may reveal a standard you have followed for years without ever writing it down.
Suppose a leader asks an agent to prepare a response to a client complaint. The agent offers a discount. The leader rejects it because the complaint concerns a broken commitment, not the price.
The correction is more than a change of wording. It explains how the leader understands the situation: acknowledge what happened, establish what can be repaired and do not use a discount to avoid accountability.
That reasoning may help with future work. But it needs to be recorded, made available when relevant and checked against what the agent does next. A correction buried in an old conversation may not carry forward.
This is where conversation becomes part of the agent’s working context.
The human explains. The agent acts. The result shows what was understood and what still needs attention. Over time, the useful lessons can become part of how the system works.
That does not mean every conversation permanently improves the agent. Keeping a correction in memory or operating instructions is also different from retraining the underlying model. The improvement has to be supported by the system and demonstrated in later work.
The human may learn something too. Trying to explain a decision can expose an assumption, an inconsistency or a better way of approaching the problem.
Understanding is not agreement
An agent that adapts to you can also become too willing to agree with you.
Naval describes this in A Return to Code. Even when he uses different models to review work, leading them toward an answer can produce agreement rather than useful challenge. He also describes agents losing context, fixing the same bug repeatedly and making quick patches when the problem needs a more substantial repair.[2]
His conclusion is direct: “you still have to guide these models.”[2]
For AGL, understanding a person’s worldview should help an agent serve their work. It should not make every belief a fact or every preference a rule that cannot be questioned.
The agent should be able to point out conflicting instructions, bring forward evidence that changes the picture and ask whether an earlier decision still applies.
Knowing you should make it more useful, not merely more agreeable.
That gives us a practical test:
Does the agent understand what matters well enough to recognise when it should question, clarify or stop?
Naval describes himself as the final gate deciding which software changes go out.[2] The same distinction matters in professional work. Preparing a proposal does not grant permission to send it. Drafting a response does not grant permission to make a commitment to the client.
The human still decides what the agent may do and remains responsible for the decisions made on their behalf.
Naval’s argument opens the conversation: speak naturally and let the technology become easier to use.
AGL’s extension is to make that conversation count beyond the immediate answer. Explain the reasons behind the work. Carry useful corrections forward. Keep room for challenge and make the limits clear.
Speak in your own words. Build the relationship.
Continue through the Professional route →
Source notes
- Naval Ravikant and Nivi, A Motorcycle for the Mind, Naval Podcast, 19 February 2026. Official transcript, particularly “The Hottest New Programming Language Is English” and “Is Traditional Software Engineering Dead?”
- Naval Ravikant and Nivi, A Return to Code, Naval Podcast, 28 April 2026. Official transcript, particularly “The Personal App Store,” “Vibe Coding Is a Video Game With Real-World Rewards,” “AI Is Eager to Please” and “Coding Agents As Customer Service Reps.”
Independent AGL editorial commentary. No affiliation, endorsement, partnership, sponsorship or technical parity is implied.