The AI Skills You Actually Need (Not the Hype)

Prompt engineering doesn't matter anymore

Early AI required precise phrasing (saying 'you are a physicist' or 'think hard about this' mattered), but modern models are robust enough that if you can give clear human-to-human instructions, you'll do fine. The technology got good enough that all the micro-optimization tricks became obsolete.

Younger people aren't naturally better at AI

The 'digital native' myth doesn't apply to AI. Junior employees often seem like they're using AI well but are actually just conduits passing along Claude's answers without understanding them. Experience and domain expertise matter more than age—you need to judge whether the output is actually good.

All three major AI models are roughly equivalent

ChatGPT, Claude, and Gemini all have similarly capable base models (ChatGPT 4.5, Claude Opus, Gemini 3.1 Pro). The real differences are in apps and harnesses (tools that let AI do things). For most people, pick one and stop worrying about which one to use.

Comparing the three major AI platforms

AI now beats humans on complex work 84% of the time

In a study of professionals with 14 years of experience (journalists, lawyers, product managers), AI completed 7-8 hour tasks in 15 minutes. Expert evaluators preferred AI outputs 84% of the time, up from 48% a year prior. Even accounting for an hour of human evaluation, you save 3x effort and cost.

AI quality improvement on complex professional tasks

The 'jagged frontier' of AI capabilities

AI is exceptionally good at some tasks and surprisingly bad at others in unpredictable ways. It struggles with jokes, writing in your specific voice, and subtle contextual nuances. Where AI is weak, human labor becomes more valuable, not less—you become the bottleneck.

AI has one voice, not yours

AI writes beautifully but with a singular, recognizable voice (slightly dramatic, loves transitions and m-dashes). It can't replicate your authentic voice for long-form work. You can improve this by giving it a large sample of your writing and having it summarize your style, then using that as a custom instruction.

Pay for the best model and use it

Spend $20/month on one of the three major platforms and actively select the latest/best model available (they default to lower-tier versions). This single choice yields huge improvement. Don't stress about which platform—just pick one and use it for real work.

Use AI for the crappy first draft, not the final product

Have AI generate a rough draft in minutes, then spend your time editing and adding your voice. This flips the traditional writing process (which spent 80% of time on the first draft). You get to do the more enjoyable work of refinement while AI handles the blank-page problem.

How AI changes the writing workflow

Tell AI to criticize you, then ask it to evaluate your thinking

AI is sycophantic by default—it agrees with you. Explicitly tell it to act as a critic. Mid-conversation, ask it to identify what you're doing wrong in your arguments, what patterns you're missing, and how you could be more persuasive. This turns it into a debate partner that strengthens your thinking.

Have AI evaluate your work from multiple personas

Ask AI to read your writing as a naive reader (what's confusing?), as a hostile expert (where would they nitpick?), as a cynic, as a critic. Giving AI personas doesn't change its ability, but it changes how it talks about your work, giving you feedback you couldn't get without consulting multiple readers.

Use AI for research and fact-checking, not just answers

All major AI models have deep research modes that are quite good. Use AI to do initial research, then fact-check its output yourself. This saves you the legwork of gathering information while keeping you in control of accuracy.

Apprenticeship as a learning model is broken

Historically, junior employees learned by doing grunt work assigned by middle managers, who evaluated their growth. Now every junior person knows less than ChatGPT and would be foolish not to use it. Every manager prefers delegating to AI over a flawed human. The feedback loop that built expertise is gone.

Learning requires effortful struggle

Neuroscience shows that learning is inherently effortful. If you shortcut through AI giving you answers, you learn nothing. This applies whether you're learning math, history, or a skill. The problem isn't AI—it's that we've become results-obsessed and efficiency-obsessed, cutting out the struggle that builds competence.

Schools and workplaces aren't built for an AI world

Education wasn't designed for a world where anyone can write essays. Work wasn't designed for instant PowerPoint generation. Law courts weren't designed for AI-generated legal briefs. Every system that relied on effort as a filter is now broken, requiring radical redesign.

When everything is generically good, taste becomes the differentiator

If Claude runs every company equally well, there's no competitive edge from quality alone. Humans who bring variation—through taste, judgment, and perspective—create competitive advantage. Your sense of taste (what you choose, who you listen to, what questions you ask) matters more than the organization delivering it.

Expertise is how you evaluate AI output

An expert can instantly spot when AI is wrong and often why—whether it's a rookie mistake you can fix or a subtle issue the AI will never grasp. Expertise becomes the bottleneck. The value of your labor shifts from producing the output to judging whether the output is good.

Jobs change shape, not disappear

Coders can write 100x more code, but if organizational processes take 2 weeks per sprint, one person becomes 100x more productive—which breaks the system. Jobs don't vanish; the balance of tasks shifts. Writing used to be 50% drafting, 50% editing; now it's 5% drafting, 95% editing. Demand for editing skill rises.

Developing taste is the new skill to teach

Taste—the ability to make choices, experience broadly, and articulate why you prefer one thing over another—is becoming more valuable than execution. Directors matter more when they can direct the entire vision. This requires a lifetime of exposure, choice-making, and vocabulary-building.

White-collar workers will lobby for protection, just like blue-collar workers did

Lawyers and doctors have guilds, associations, and congressional representation. AI is very good at legal and medical work, but these professions won't go quietly. Expect laws requiring human sign-off even if AI is better. Coders lack this protection and will face more disruption.

80% of jobs in 20 years don't exist yet

Technologists often cite that 80% of jobs today didn't exist 20 years ago, implying new jobs will emerge. This is true but doesn't help people living through the transition. The real question is how to cushion people during the chaos, not whether jobs will eventually exist.

We have agency at two levels

Societal agency: policy-making, organizing, writing to Congress still works. Individual agency: using AI to augment your team, showing positive examples of humans thriving with AI, building tools that expand human capability rather than replace it. The default plan (fire everyone, replace with AI, raise profits) is dangerous; the alternative is showing how augmentation works.

The real fear: chaos during transition, not apocalypse

Even if AI works out fine long-term, living through the Industrial Revolution was miserable (see Dickens). We're facing a period of chaos with haves and have-nots, social change, and ripple effects (deep fakes, information trust, job displacement). Policy needs to address this complexity, not pretend it's either all good or all bad.

AI tutoring is the pedagogical holy grail

Personalized education—where a tutor adapts to your level, challenges you at the right difficulty, and makes topics relevant to your interests—was impossible at scale until now. AI tutors can do this. This is a genuine positive that could transform learning, not just a threat.

In-class assessment and active learning are the solution

Schools solved this with calculators: use AI for some tasks, not others. Do in-class testing, active learning, and discussions. Have students use AI tutors that ask questions rather than give answers. Make them launch projects in areas they have expertise. The solution exists; it requires changing how we teach.

Technology changes what our brains do, not if they work

We stopped memorizing phone numbers when we got digital diaries. We stopped doing matrix multiplication by hand when calculators arrived. We gave up cursive. Each time, we chose what to give up. AI will change what thinking looks like, but that's not the same as destroying thinking. The choice is ours.

AI privacy is similar to Gmail privacy

All three platforms offer opt-out of training on your data if you pay. Your chat history isn't searchable by others (no one can query ChatGPT to read your emails). It's enterprise software with the same risks as Gmail—the company has your data, but it's not publicly accessible. The real risk is if AI gets access to your computer and browser.

The real security concern is computer access

If you give Claude or Codex access to your computer, files, email, and web browser, the theoretical risk is someone convincing the AI to send them your money or data. This hasn't happened yet but isn't impossible. This is a different category of risk than data privacy.