SkillCurio Dispatch · Issue #2
AI skills are changing. Here’s what matters now.
The valuable AI skill is increasingly becoming the ability to direct, evaluate, integrate, and supervise AI—not simply use it.
AI skills are changing.
Not because prompting suddenly doesn't matter, or because everyone needs to become an AI engineer.
Something more interesting is happening: the valuable skill is increasingly becoming the ability to direct, evaluate, integrate, and supervise AI—not simply use it.
Several developments this month point in that direction.
What changed
1. AI literacy is becoming AI supervision
Boston Consulting Group's latest global AI-at-Work research surveyed 11,749 workers across 14 markets.
72% said AI has considerably changed the skills expected in their roles.
Even more interesting: 47% said they now spend more time managing and directing AI than doing the underlying work itself. Among frontline employees, regular AI use has reached 74%.
That changes what “AI literacy” means. Knowing how to write a good prompt is useful. But workers increasingly need to know how to define the task, give AI the right context, evaluate what comes back, recognize when the system is wrong, and decide when human judgment needs to take over.
2. A new kind of AI role is appearing between technology and the business
Take a look at what McKinsey is asking of its Forward Deployed Engineers.
The role involves building and deploying AI systems, but also integrating them with enterprise systems, evaluating models, monitoring performance, meeting security requirements, training users, and translating customer problems into technical implementations.
That distinction matters. AI can reduce some of the effort required to produce software. But that can make understanding the environment where the software has to work even more important.
If you've spent years in IT, operations, security, support, finance, healthcare, law, manufacturing, or another domain, don't automatically assume AI means starting your career over.
Your existing knowledge of systems, users, workflows, security, troubleshooting, and business processes can become part of your AI skill set.
The goal isn't necessarily:
Old career → AI engineer
It may be:
Existing expertise + AI capability → new role
3. The practical AI skill stack is becoming clearer
A September analysis of 1,056 live U.S. AI Engineer openings looked at what employers actually mention in their job postings.
Among the 707 postings whose full descriptions were analyzed, Python appeared in 68.4%, LLMs in 52.8%, RAG in 34.7%, AWS in 27.2%, prompt engineering in 19.7%, CI/CD and REST APIs in 16.7%, Azure in 15.9%, Kubernetes in 14.9%, observability in 13.3%, vector databases in 12.7%, and Docker in 11.5%.
There is an important caveat: these are mentions, not necessarily requirements, and the analyzed postings aren't a random sample of the entire job market.
But the pattern is useful. Prompt engineering hasn't disappeared. It's becoming one component of a larger systems skill set.
For someone moving beyond basic AI use, a practical progression might look like:
AI fundamentals → Python/API basics → RAG → evaluation → deployment → observability and governance
You don't need to learn all of that this weekend. You need to know what comes next.
4. Human expertise may become more valuable as AI improves
Today's news from Snorkel AI is worth watching for a reason that has little to do with venture capital.
Snorkel raised $350 million at a $3.5 billion valuation as demand grows for sophisticated training data and reinforcement-learning environments. Its annualized revenue run rate has reportedly grown from roughly $20 million to more than $350 million in about a year.
But here's the part that caught my attention: Snorkel combines AI automation with specialists in fields including coding, law, and medicine to design scenarios, tasks, grading rubrics, and high-quality data for advanced AI systems.
As models become more capable, valuable training and evaluation problems can become more specialized. AI can generate enormous quantities of generic content. Expert judgment is harder to manufacture.
That suggests an important career principle: Don't throw away the expertise you already have. Learn how to make it useful to AI systems.
Your next learning step
Build one tiny RAG system
If you're already comfortable using ChatGPT, Claude, Gemini, or another AI assistant, don't spend the next two weeks collecting more prompts. Build something small.
Choose 3–5 documents you understand well: product documentation, technical notes, company policies, research material, or documents you've written yourself.
Then use an AI development tool or tutorial to create a simple retrieval-augmented generation experiment. Your goal isn't production software. Your goal is to understand this:
documents → chunks → retrieval → model → answer → evaluation
Then ask questions that your documents cannot answer. Does the system admit that the information isn't available? Does it retrieve irrelevant material? Does it confidently invent an answer?
Now you're learning something much more valuable than another prompt trick: the difference between a convincing AI answer and a grounded AI answer.
Time budget: 60–90 minutes.
Success condition: You can explain why your system retrieved a particular piece of information—and recognize when it should say, “I don't know.”
The SkillCurio Signal
What if the future of AI education looks less like school—and more like work?
One of the more interesting education experiments emerging right now is the proposed Horowitz Andreessen Academy.
The technology-backed program is being designed as an alternative to the traditional four-year college experience, with plans for students to spend as much as 80% of their time on hands-on learning and internships.
The program is still experimental and unaccredited, so whether the model ultimately works remains to be seen. But the idea behind it is worth paying attention to, because we're seeing versions of the same philosophy elsewhere in AI education.
Microsoft's recent Agent-a-Thon wasn't simply about teaching people what AI agents are. Participants moved through different tracks where they actually built agents—from introductory no-code projects to more advanced enterprise workflows.
OpenAI Academy's HBCU Innovation Summit similarly moved participants from instruction into a Codex Build Challenge, where teams created projects addressing real problems in campuses, workplaces, and communities.
Different organizations. Different audiences. But a remarkably similar learning model is emerging:
Learn → build → test → explain → improve
Instead of:
Watch → memorize → certify
That doesn't mean traditional education or credentials are becoming irrelevant. Some professions will continue to require them, and foundational education still matters.
But AI may increase the value of something credentials don't always demonstrate: Can you actually apply what you've learned?
For someone trying to reposition their career around AI, that leads to a practical question: What can you build, automate, analyze, or improve that demonstrates your new capability?
A small working project may tell an employer—or even yourself—more about your progress than another completed video course.
One thing to remember
You don't need to learn everything happening in AI.
You need to become better at identifying which AI capabilities complement what you already know.
Start there.
Then build something.
Originally published in SkillCurio Dispatch on September 22, 2026. This web edition preserves the newsletter text and source links, with additional SkillCurio learning-path suggestions.