From AI burnout to AI native: the 5-level blueprint to actually using agents.

So, earlier this week we asked you (our readers) what you wanted more of. The #1 answer, by a landslide: “Stop telling me what tools exist. Show me how to USE them.”

Fair. So we’ve spent all week watching every practical AI tutorial we could find, and one framework in particular stood out. We think this framework is particularly helpful, especially for novice AI users, but it’s helpful for triaging where you are at in your AI adoption journey no matter what level you are at below.

First up, the TL;DR:

• The Productivity Paradox: Developers are building full software suites in hours, but early adopters are reporting severe burnout from the relentless pace of AI tools.
• The "How" is Dead: AI is collapsing the execution phase of work. The new premium skills are taste, judgment, agency, and curiosity—knowing what to wish for.
• The Agent Era: Tools like Claude Cowork have moved AI from the cloud onto your local desktop, allowing agents to access files, control browsers, and execute "Skills" in parallel.
• Model Specialization: Testing shows you shouldn't rely on just one model. You need creative models (like Opus 4.6) for building, and literal models (like Codex GPT-5.3) for reviewing and catching bugs.

Now here’s the framework in brief: Product leader Peter Yang broke AI adoption into five levels:

• Level 1: Asking AI questions instead of Google. Peter says that's where 99% of people are stuck. It's fine, but it's barely scratching the surface.
• Level 2: AI for daily work. This is where 80% of the value lives, and it only takes three things: a voice dictation app (like Wispr Flow), a meeting notes tool (like Granola), and Claude / ChatGPT Projects with custom context for recurring tasks. Peter even helped his 60+ year-old parents set this up.
• • P.S: To write this newsletter every day, this is basically all we use as well.
• Level 3: Prototyping before you spec. Instead of writing a 10-page document nobody reads, clone your product's UI by building a working version in Claude Artifacts, Lovable, or Google AI Studio and iterate on something people can actually touch. As Peter puts it: code is now easier to create than a slide deck.
• • Not an app developer? Same principles apply:
  • • A marketer builds a working landing page instead of briefing one.
    • A finance analyst builds an interactive dashboard instead of a PowerPoint about the dashboard.
    • A consultant hands a client a clickable demo instead of a 40-slide deck.
  • The core idea: stop describing things and start showing them. People give better feedback on something they can touch. The prompt is where you share your raw idea; the “artifact” (or the thing you create with AI’s help) is where you share your refined ideas with other humans.
• Level 4: Building full apps. Peter's built 14+ apps with AI, from retro space shooters to board game finders, often while watching Netflix. His advice: tell AI your problems, create a plan with three milestones, and think of yourself as the manager giving feedback to a very capable intern.
• • For non-devs, think about the tedious thing you do every week that makes you mutter “there should be an app for this.”
  • • A recruiter builds a candidate scoring tool.
    • A teacher builds a quiz generator from their own lesson plans.
    • An event planner builds an RSVP tracker with auto-reminders.
  • These used to require hiring a developer. Now you describe the problem in plain English and iterate.
  • The two easiest implementations:
  • • 1. a simple automation, either in Codex or with Tasklet.
    • 2. a mini-app, like a chrome extension you can host in your browser, a private app in AI Studio, or something you can host locally (just ask Claude Code / Cowork or Codex to help you do that).
• Level 5: Personal AI agents. This is the frontier. People like Nat Eliason are building autonomous bots that run entire businesses (his made $3,500 in its first week selling a single PDF). The interface of the future, Peter argues, is lying in bed texting your AI agent vague instructions.

So where do you start? As writer Alberto Romero put it perfectly: you don't need to "master" AI. You need competence: enough skill to get real value without wasting hours on dead ends. The good news is, if you’re starting today, the AI models are as good as they’ve ever been. It’ll be easier to get value out of AI right away starting today than any time that came before.

The best part? Alberto thinks you can go from zero to competent in 8 hours. Peter says you can reach Level 4 in a single week (how about you start today?). But here’s the thing: don't jump straight to Level 5. The compounding value is in the earlier levels, and they're shockingly easy to set up.

Why this matters: Romero nailed it here as well: AI is collapsing the "how" so fast that the bottleneck is now the "what." What do you want to build? What problem do you want to solve? As he puts it, AI is the closest thing in the world to a genie lamp, and most people are walking around with it in their pocket, still spending 99% of their effort on execution instead of deciding what to wish for.

Now let's dive into all of that with more depth below.

Two Viral Posts Just Told You the Future of Your Career. They Disagree on Everything Except One Thing.

This week, AI founder Matt Shumer published a 5,000 word essay that went nuclear: "Something Big Is Happening." His claim? After GPT-5.3 Codex and Opus 4.6 dropped on the same day, he no longer needs to do the technical work of his own job. He describes what he wants in plain English, walks away for four hours, and comes back to finished software. No corrections needed.

He backs this up with data from METR, which tracks how long AI can work independently. A year ago, it was about 10 minutes; today, AI handles tasks that would take a human expert five hours. That capability doubles roughly every seven months—and might be accelerating to every four. Even wilder, GPT-5.3 Codex's documentation explicitly notes it was "the first model that was instrumental in creating itself".

The piece reads like a survival guide for your career. Shumer (an AI startup CEO of six years) says the latest models don't just execute instructions; they show something that feels like judgment. Like taste. The kind of intuition people always said machines would never have.

His advice is blunt: spend one hour a day experimenting with AI. If you do that for six months, you'll understand what's coming better than 99% of the people around you. The bar is on the floor, he says.

The exact same day, veteran engineer Steve Yegge (40 years in tech across Amazon and Google) published "The AI Vampire", a stark warning that this 10x productivity boost is draining people dry. He describes experiencing sudden "nap attacks" after long coding sessions with AI, comparing the tool to a slot machine that doles out relentless dopamine shots.

While Yegge agrees the 10x productivity boost is real, he argues we're ignoring the cost. People are burning out. Companies are designed to extract every ounce of that new productivity, leaving workers drained with nothing to show for it.

His formula is elegantly simple: $/hr. You can't always control your salary. But you control how many hours you give away.

In fact, want to guess what his proposed fix is? We need to shrink the workday to 3-4 hours. Not because people are lazy, but because AI concentrates all the hard cognitive work (decisions, judgment calls, problem-solving) into a compressed window that's exhausting to sustain. I mean, on principle, I don't disagree.

Now, the crux of his argument is that we're stuck between two extremes. If you're 10x more productive, who actually benefits? Not the employee. You're just working 8 hours at 10x speed, and your employer captures all the value while you burn out; now, if you only work 1 hour at 10x speed, then you capture 100% of the value but your company eventually gets crushed by competitors who move faster (and, well, work more hours).

Here's our take:

Both are right, and both are incomplete. The real takeaway isn't "panic" or "slow down." It's this: going slow doesn't guarantee your job. Going fast doesn't either. Doing things well AND fast? That's your job security now.

Neither speed alone nor quality alone will protect your career. You need both. And getting there requires curiosity, humility, and a willingness to rethink how you work. As Deedy Das put it, just look at everything AI has unlocked in 10 years: Hollywood-grade video, senior-level code, protein prediction, competition-level math. If you're still saying "95% of AI pilots fail," you're probably using the wrong product.

IMO, the economy is in the slow-but-speeding-up process of splitting into two lanes: there's work where humans create for other humans (creators, leaders, salespeople, relationship-builders), and there's work where machines handle the corporate machinery. If your role is a component in a machine, that component will inevitably get swapped. If you're a human making things for other humans, you'll be fine... so long as the corporation you work at doesn't think of you like a component.

To find the middle ground between these two extremes—to actually use these tools without succumbing to the vampire—you need a system. We already wrote about compound engineering, which makes a ton of sense to us. But it's not quite right... it's not perfect on it's own, especially not for non-technical people. The ideas are good, but we wanted something more practical.

Well, product leader Peter Yang recently broke AI adoption into five clear levels. It’s the clearest roadmap we’ve seen to transition from a burnt-out operator to an AI native. Let's dive into it. Using tools like v0 (by Vercel), Bolt.new, or Google AI Studio, you can paste a screenshot of your existing UI and ask the AI to add a new feature. In a live demo, Peter cloned the YouTube Studio dashboard and replaced a panel with AI content suggestions in 20 seconds. "Code is now easier to create than any kind of document, slide deck, or design," he notes.

Level 1: Everyday Answers (Where 99% of People Are)

At this stage, you're just asking AI questions instead of Googling them. Peter notes he recently fixed his toilet by sending ChatGPT photos and questions. It’s a good start, but it’s barely scratching the surface.

Writer Alberto Romero's guide, "Learn to Use AI Competently in 1 Day", highlights that you don't need to "master" AI—you just need basic competence. His golden rule is just six words long: be specific about what you want. Don't type "write me an email"; include the goals, constraints, and edge cases you’d give a human colleague.

Level 2: AI for Daily Work (Where 80% of the Value Lives)

According to Peter, Level 2 accounts for 80% of the value he gets from AI. You only need three things to unlock this:

• Voice dictation: Apps like Wispr Flow, Monologue, or Super Whisper let you speak instead of type. AI is incredible at parsing stream-of-consciousness thought.
• Meeting notes: Tools like Granola run quietly in the background, allowing you to ask questions mid-call like, "what have people been talking about?" without adding annoying bots to the calendar.
• Custom Context: Set up dedicated Claude or ChatGPT Projects loaded with your personal context. Peter set this up for his 60+ year-old parents to monitor their health and plan vacations.

Enter Claude Cowork: This is also where Anthropic's new desktop application, Claude Cowork, shines. As Ben AI points out, Anthropic now has three tiers: Claude Chat (for brainstorming), Claude Code (for app building), and Claude Cowork, which is an active "doer" for day-to-day office tasks.

Cowork doesn't run in the browser; it's a desktop app that directly accesses your files and tools. Boris (via Greg Isenberg) demonstrates its power with a "receipts hack": dump poorly named PDFs into a folder, and Cowork will automatically open them, read the dates and vendors, rename the files, and even extract the data into a Google Sheet. If it lacks a direct API connector, it has a "browser fallback" where it literally opens a window and scrolls through X or Google to do background research for you.

Level 3: Prototyping Before You Spec

This level flips the traditional corporate workflow on its head.

Instead of writing a 10-page spec document, handing it to a designer, and then an engineer, you just build a working prototype. Using tools like v0 (by Vercel), Bolt.new, or Google AI Studio, you can paste a screenshot of your existing UI and ask the AI to add a new feature. In a live demo, Peter cloned the YouTube Studio dashboard and replaced a panel with AI content suggestions in 20 seconds. "Code is now easier to create than any kind of document, slide deck, or design," he notes.. In a live demo, Peter cloned the YouTube Studio dashboard and replaced a panel with AI content suggestions in 20 seconds. "Code is now easier to create than any kind of document, slide deck, or design," he notes.

The philosophy shift: Why is this so profound? Alberto Romero's second essay, "You Spent Your Whole Life Getting Good at the Wrong Thing", explains that AI is acting like a genie lamp, instantly handling the "how" (the execution). Because the "how" was always so expensive and difficult, we spent our whole lives optimizing for it. Now, the bottleneck is entirely the "what". Your career now depends on taste (selection), judgment (evaluation), agency (initiation), and curiosity.

Level 4: Building Full Apps

Once you know what you want to build, you enter Level 4, where Peter has built over 14 apps (like retro space shooters) often while just watching Netflix. His advice: give the AI your problems, ask it to create a 3-milestone plan, and treat yourself as the manager reviewing an intern's work.

But which model should you use? Product engineer Claire Vo recently tested OpenAI's Codex GPT-5.3 against Anthropic's Opus 4.6 on a massive production codebase.

• Codex (GPT-5.3): She found it highly literal. When asked for an enterprise site, it literally wrote "if you're here for product-led growth, click here" in the copy.
• Opus 4.6: Much better at creative, open-ended planning. It produced a gorgeous, on-brand redesign almost immediately.

The winning workflow is combining them: Opus builds the code (because it's creative), and Codex reviews the code (because it's ruthlessly literal and catches every bug). Using this dual-model strategy, Claire merged 44 pull requests, 98 commits, and 93,000 lines of code across 1,088 files in just five days. (Note: Opus 4.6 is 6x the price of standard models, so use it strategically!)

Level 5: AI Agent Teams (The Frontier)

The ultimate interface of the future is lying in bed, texting your AI vague instructions, and letting it handle the rest. Peter Yang is already seeing this: entrepreneurs are building custom "LFG" commands that automate two hours of engineering work, and he says Nat Eliason recently gave an AI its own X account, Stripe, and GitHub keys—and it actually made $3,500 selling a PDF autonomously in one week.

So how do you achieve this? You need Skills and Agent Teams.

The power of Cowork Skills: Ben AI highlights that Cowork "Skills" are the evolution of Projects. Instead of hard-coding automations in Zapier, you manually walk Claude through a workflow once (e.g., download a YouTube transcript, write a newsletter, format the output). When it's done, you simply say, "Save this process as a Skill". Now, you can run multiple Skills in parallel within the same chat window.

Claude Agent Teams: Developer Leon van Zyl recently tested Anthropic's new experimental Agent Teams feature. He built a fitness app using five distinct agents operating simultaneously: a UX designer, a backend dev, a technical architect, a database expert, and a "devil's advocate".

Unlike traditional sub-agents that each one works alone and reports back to a main agent, Agent Teams use direct peer-to-peer messaging. They sit in the same virtual room, view the same task list, and talk to each other directly (so the API agent knows exactly what the database agent is doing). Add a simple property to your settings.json file, and you have an entire autonomous IT department on your laptop.

Now here's a tactical setup guide:

Before you run off to build an army of autonomous agents, let's get practical. Here is the exact, step-by-step cheat sheet to actually set these tools up today:

• 1. The Claude.md memory hack: Create a file called Claude.md and check it into your code repository. Every time Claude makes a mistake, add the correction to this file. Claude reads it before every session, creating a "never repeat feedback" loop.
• 2. How to ACTUALLY install "Skills": Go to Settings > Capabilities > Add > Upload a skill. Download pre-built skills from places likesmithery.ai/skills. Attach your business context files (like your ICP) to make it your own.
• 3. The manual connector hack (MCP): If an app isn't natively supported, Google "[your software] MCP Claude". Open the Claude desktop config.json file, paste in the MCP configuration code, and the connection is live.
• 4. The app-building prompt: If you don't know what to build, ask: "What are some simple apps we can build to help me take time back? Here's context about my life". Always ask for a robust plan with three clear milestones before writing code. Use Replit for beginners, Cursor for more control, and Claude Codei n the terminal for maximum power.
• 5. The "devil's advocate" setup: Add CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS to your settings.json file to enable Agent Teams. In your prompt, write: "Create an agent team to explore this from different angles". Always assign one agent the role of "devil's advocate" to question the others—it dramatically improves the output.

Armed with these setups, you have everything you need to start climbing the ladder.

The bottom line is this: Peter says you don't need to jump straight to Level 5. As Peter Yang notes, you can easily scale from Level 1 to Level 4 in a week just by adopting voice dictation, meeting notes, and prototyping. The tools are officially here, and the "how" has been solved. The only question left is: What are you going to wish for?