Episode 9 / Feb 28, 2026

Is AI productivity as simple as using more tokens?

How does Peter Steinberger spend $20k/month on tokens, and why? Based on their own experiments, Eric and John talk explain why autonomous loops are the next productivity frontier for AI.

Take it with you

Listen on your platform

Spotify ↗Apple Podcasts ↗Substack ↗

Listen here

The conversation in writing

Show notes

Summary

Eric and John trace the rapid evolution of AI productivity, from prompt engineering to context engineering to autonomous loops. They land on a surprising insight: the biggest unlock isn't how you talk to AI, it's how much you let it run without you. They use OpenClaw's heartbeat file, real token-cost math, and the concept of long-horizon planning to argue that the bottleneck is shifting from prompt engineering skill to outcome definition and, ultimately, to human adoption speed.

Key Takeaways

Prompt engineering is already productized: tools like v0’s prompt enhancer and Claude's plan mode have absorbed what used to be a manual skill.
The real token spend comes from autonomy, not interaction: running multiple agents on loops is how you get to $15–20K/month, not by typing faster.
Define the outcome, not the process: autonomous loops work best when the destination is crisp; vague goals still need human-in-the-loop collaboration.
Long-horizon planning is the emerging skill: if AI compresses three years of execution into a quarter, you need to plan at a level of detail nobody's practiced.
User adoption is the true ceiling: even if you can ship three years of product in three months, humans can't consume it that fast, so the bottleneck moves from build to adoption.
Get (tokens) while the getting's good: $200/month subscriptions currently deliver thousands in real token value, but that arbitrage won't last forever.

Notable mentions and links

Agent skills are reusable capabilities for AI agents that you can manually install. They are mentioned as part of the progression from prompt engineering to context engineering and beyond.
Claude's plan mode (and similar features in other tools) are framed as productized versions of prompt engineering. Boris, the inventor of Claude Code, explained on Lenny's Podcast that plan mode is just a prompt telling the model to plan and not write code.
The heartbeat file is an OpenClaw text file with instructions that a scheduled job reads every 30 minutes. The AI agent wakes up, executes tasks autonomously, then goes back to sleep.
Anthropic's agent experiments, like building a C compiler, are cited as examples where clearly defined outcomes make autonomous loops viable.

Every word of the conversation

Transcript

Find a word or phrase

00:00:00,080 --> 00:00:30,360 [Eric]

[upbeat music] Welcome back to the Token Intelligence Show, where John and I are trying to figure out AI and life, and you get to follow along with our successes and failures. We have a great topic today. John, I'm amazed at how the productivity hacks with AI

00:00:30,360 --> 00:00:45,320 [Eric]

have changed so unbelievably fast over the last couple years. And I think, correct me if I'm wrong, but I think the very first iteration of this was prompt engineering.

00:00:45,320 --> 00:00:46,080 [John]

Yes.

00:00:46,080 --> 00:01:01,600 [Eric]

And it eventually became ca- it was eventually called prompt engineering. But even before that, I remember it was like in a past episode, we mentioned sort of, uh, hacks for Google search, where you could do site colon and it would-

00:01:01,600 --> 00:01:01,610 [John]

Yeah

00:01:01,610 --> 00:01:03,100 [Eric]

... you know, Google would search a particular site or whatever.

00:01:03,100 --> 00:01:03,400 [John]

Right. Right.

00:01:03,400 --> 00:01:07,460 [Eric]

And so the AI version of this was

Related conversation / Episode 8

Navigating skill atrophy in the AI age

Eric stopped using AI for personal writing. Why? As you outsource to AI, you must decide which skills to keep sharp. Hand-coding is fading, but thinking, storytelling, and taste are timeless.

Explore this episode ↗

Token Intelligence

The conversation

Summary

Key Takeaways

Prompt engineering is already productized: tools like v0’s prompt enhancer and Claude's plan mode have absorbed what used to be a manual skill.

The real token spend comes from autonomy, not interaction: running multiple agents on loops is how you get to $15–20K/month, not by typing faster.

Define the outcome, not the process: autonomous loops work best when the destination is crisp; vague goals still need human-in-the-loop collaboration.

Long-horizon planning is the emerging skill: if AI compresses three years of execution into a quarter, you need to plan at a level of detail nobody's practiced.

User adoption is the true ceiling: even if you can ship three years of product in three months, humans can't consume it that fast, so the bottleneck moves from build to adoption.

Get (tokens) while the getting's good: $200/month subscriptions currently deliver thousands in real token value, but that arbitrage won't last forever.

Notable mentions and links

Agent skills are reusable capabilities for AI agents that you can manually install. They are mentioned as part of the progression from prompt engineering to context engineering and beyond.

Claude's plan mode (and similar features in other tools) are framed as productized versions of prompt engineering. Boris, the inventor of Claude Code, explained on Lenny's Podcast that plan mode is just a prompt telling the model to plan and not write code.

The heartbeat file is an OpenClaw text file with instructions that a scheduled job reads every 30 minutes. The AI agent wakes up, executes tasks autonomously, then goes back to sleep.

Anthropic's agent experiments, like building a C compiler, are cited as examples where clearly defined outcomes make autonomous loops viable.