Up until the beginning of 2026, everyone who wasn't an engineer engineer and understood just enough about AI to feel comfortably using it wanted to learn prompt engineering. People took courses on it. Companies hired for it. LinkedIn influencers built entire brands around "the perfect prompt."
Well, it's my belief that in 2026, the best AI users don't write elaborate prompts anymore. In fact, they barely think about prompting at all. The meme is that something changed in December 2026, and most people haven't fully caught up to speed that perfect prompts are dead. Here's what replaced it.
First up, the TL;DR
Two years ago, "prompt engineering" was the hottest skill in tech. People took courses, companies hired for it, and LinkedIn was drowning in "10 prompts that will change your life" posts.
Here's what happened: the models got good enough that you don't need to trick them anymore.
In 2023, you needed weird hacks to get useful output. "Pretend you're an expert." "Think step by step." "I'll tip you $200." In 2026, you just... talk to them. The skill moved upstream.
What replaced prompting is something educators have taught for decades: metacognition (thinking about your own thinking). The best AI users today are doing three things:
- Thinking clearly before they ask. If you can't explain what you want to a smart coworker, you can't explain it to AI. This sounds obvious. It's the hardest part.
- Building context systems. Anthropic just revealed their engineers have hundreds of "Skills" (structured folders of instructions, scripts, and examples) that give Claude the context it needs before a conversation even starts. The difference between asking AI a question cold vs. inside a system that already knows your standards is enormous. Most people don't know this layer exists.
- Exercising editorial taste. Anthropic engineer Thariq Shuja spent 1.5 weeks writing a post about Skills. Claude did research and a first draft that helped structure his thinking, but he threw out basically all the copy. His takeaway: "You have nothing until you have everything." The AI accelerated his process. His judgment shaped the final product.
These skills share something interesting: clear communication, system design, quality judgment. They're the same skills that made people effective before AI existed. AI just amplifies them. Someone who thinks clearly gets 10x more value from these tools than someone memorizing prompt formulas.
AI fluency looks less like coding and more like management. You're directing a capable collaborator. The people who are best at this aren't prompt engineers. They're clear thinkers who happen to use AI.
If you're an educator, a manager, or anyone whose job involves thinking (so... everyone), that's actually encouraging. The most valuable AI skill is the one you've been building your whole career. Your decision making.
Now, in case you don't believe me, let me make my case abundantly clear.
The Prompt Engineering Era: 2022 to 2024
The prompt engineering window lasted roughly two years. During that time, AI models were powerful but finicky. Getting good output required specific techniques: structured instructions, role assignments, chain-of-thought triggers, even emotional manipulation ("this is very important to my career").
These worked because early models needed explicit scaffolding to activate their full capabilities. "Think step by step" genuinely improved reasoning quality. "You are an expert in X" helped the model retrieve the right patterns from its training data.
Then the models improved. Claude, GPT-4o, Gemini, and their successors got better at interpreting intent from natural language. The scaffolding on the individual prompt level that used to be mandatory became optional. "Think step by step" still helps occasionally, but modern models reason through complex problems without being told to (at least when you turn "thinking" or "reasoning" on).
Most AI advice you might read as a total beginner hasn't caught up to this change. The internet is still full of prompt templates and cheat sheets designed for a generation of models that no longer needs them. It's like teaching someone to double-clutch a car with an automatic transmission. The underlying mechanics are interesting, but the practical skill has moved on.
What Metacognition Looks Like in Practice
Metacognition is an academic word for something simple: being aware of how you think. Educators use it constantly. When a teacher asks a student "what strategy are you using to solve this?" or "how do you know your answer is right?", that's metacognition in action.
Applied to AI, it shows up in three layers. Each one is more powerful than the last.
Layer 1: Knowing what you actually need.
This is where most people stall. They open an AI tool and type something vague, then get frustrated when the output is vague back.
A concrete example: "Help me write a marketing email" produces generic output because it's a generic request. The fix is spending 30 seconds thinking: Who's receiving this? What do I want them to do? What tone fits my brand? That 30 seconds of self-reflection transforms the interaction entirely.
Metacognition here means thinking about your thinking before you externalize it. And it's a skill you can practice. Every time you catch yourself typing a vague request and pause to clarify your own intent first, you're building the muscle.
Layer 2: Building systems that hold your context.
This is the biggest gap between casual and power users, and a recent post from Thariq Shuja of Anthropic about how the company uses "Skills" is the best illustration we've seen:
- Anthropic's engineers have hundreds of Skills in active use.
- These "Skills" are structured folders of instructions, scripts, reference code, examples, or configuration files that give Claude Code the context it needs before a conversation starts.
- Some examples:
- A frontend design skill teaches Claude their specific design system and common mistakes to avoid.
- A data-fetching skill tells Claude which database tables to join and where the canonical user ID lives.
- A code review skill spawns a separate Claude instance with fresh eyes to critique work.
The key insight on skills from Thariq: the highest-value content in any Skill is the "Gotchas" section, built from real failures. I.e., the specific things Claude (or if you're using ChatGPT, Codex) gets wrong when using your tools. This is pure metacognition: you have to understand your own workflow, your own standards, and your own failure modes well enough to externalize them into a system.
A lot of times this means you have to try to do it wrong first before you get it right.
Anthropic's Skills docs break Skills into two types that map cleanly to how you'd use them:
- Reference content, as in your knowledge, conventions, style guides, and domain expertise that Claude applies to whatever you're working on.
- Most people should start with reference content. Write down what you know that the AI doesn't.
- Task content (step-by-step instructions for specific actions, like "deploy this" or "review this PR").
- The task content comes later, once you've watched the AI work enough to know which processes are worth automating.
Now, before you get intimidated, you don't need to be an engineer to apply this principle. A teacher could build a custom instruction set that captures their grading rubric, their students' common misconceptions, and their preferred feedback style. A marketing manager could create a project workspace with brand guidelines, past campaign performance data, and their company's tone of voice. The concept is the same: invest time understanding your own process (as in, try to do it with AI, have the AI fail, then write down things the AI should do differently), then encode that understanding into context the AI can use.
Thariq's advice for building good Skills applies to any context system: "Don't state the obvious. Focus on information that pushes the AI out of its default patterns." In other words, the value comes from your specific knowledge, the things only YOU know about how your work actually gets done.
Layer 3: Taste and iteration.
This is the editorial judgment layer, and Thariq's own writing process is the perfect example.
He used Claude to catalog and categorize all of Anthropic's internal Skills, generate examples, and produce a first draft of his blog post. Then he threw out basically all the copy Claude wrote. Claude essentially accelerated his research and helped structure his thinking. But it was his taste and judgment shaped the final product. The result: 5.5 million views in two days. Not bad!
This mirrors how a good manager works with a direct report. Spoiler alert: they don't write the report themselves. They don't give instructions so detailed they might as well have done it. They review the work, spot what's off, and give targeted feedback: "The second section buries the lead. Tighten the opening. Fix that weird chart crime on page six." Effective AI use works exactly the same way. You evaluate, redirect, and refine. The skill is in the back-and-forth. And then the back-and-forth becomes the Skill (with a capital S).
For more on Skills, check out Anthropic's docs around them, and their new free course teaching you how to use them.
Why This Scales
A prompt template gives you one good output for one specific task. You really wanna spend the rest of your career writing one off prompts? That is NOT productive, and it certainly isn't "automating" your work. Instead, Skills (and the process to create them, through metacognition and thinking about what you need to communicate to your digital coworker to get it to help you) give you better output across every task, with every tool, indefinitely. This is true of Skills, and it's also true of this explore, refine, and exploit loop in the meta-sense.
Someone who learns to think clearly about what they need and how to get it can switch from Claude to ChatGPT to Gemini to whatever launches next year. The skill transfers because it lives in you, not the tool. And when you take the time to turn these skills into Skills (sorry I keep doing this; it's important to mention both, and also, it's funny to me), you can quite literally transfer the skill from one tool anywhere else; they become pre-packaged memories of how to do things you can call at a moment's notice. You can think of them like procedural memory; once you learn how to ride a bike, you've pretty much got it.
At a technical level, Skills follow an open standard called Agent Skills that works across multiple AI tools. A Skill you build for Claude Code today can work in Cursor, Codex, Gemini CLI, or whatever coding agent ships next quarter. You're building portable expertise you can take to any platform.
This also explains why some people find AI transformative while others find it underwhelming or flat out busted. It's rarely about which model the person is actually using (although that does make a difference). The person who spends 30 seconds clarifying their own thinking before typing (or brain dumping as they type, which also works) gets dramatically better results than the person who memorized or copy-pasted 50 prompt templates, but then gets stuck the next time they run into something they don't have a saved prompt for.
For educators specifically (and we know many of you reading this are): this maps directly to frameworks you already teach. Bloom's taxonomy, self-regulated learning, reflective practice. AI fluency is an application of metacognitive skills to a new medium. You're not behind. You might be ahead of the curve. You just need to update your AI skills for the Skills era. I don't think it's bombastic to call this a new paradigm; it clearly is as everyone has rushed to adopt it, and it's one of the main things that makes the open agent system OpenClaw so viral.
What You Can Do Today
Four things, in order of effort
- 30 seconds: Before your next AI conversation, write down (even mentally, or spoken out loud using the voice transcribe tool) what you actually want. The specific outcome, the audience, what "good" looks like. That clarity is the real prompt. Always has been.
- 30 minutes: Set up a custom instruction in your AI tool of choice that captures your role, your preferences, and your standards for your most common tasks. This is Layer 2 in miniature. You're building your first piece of context infrastructure, even if it's just a paragraph.
- A weekend project: Turn that custom instruction into an actual Skill. A Skill is just a folder with a SKILL.md file: some YAML frontmatter (a small block of metadata that tells the AI when to use the Skill) and markdown instructions underneath. That's it. You don't need to be an engineer. If you can write a clear set of instructions for a colleague, you can write a Skill.
- Start with the task you repeat most often. Write down what you always have to re-explain to the AI. Add a Gotchas section with the mistakes it keeps making. Congratulations: you just externalized your metacognition into a reusable system. That's the whole game.
- Personal Skills stored in your user folder are available across every project you work on. Build it once for the task you repeat most, and it follows you everywhere. Project-level Skills can live in a shared folder so your whole team gets them automatically.
- As mentioned earlier, Anthropic just launched a free course on Agent Skills that walks you through building them from scratch, and their public Skills repo on GitHub has templates and examples ranging from document processing to creative design to enterprise workflows. Browse those to steal structures (Principle #7: copy well, not blindly), then make your own.
- Ongoing: Practice the follow-up. The first output is almost never the final product. Learn to evaluate: what's right, what's off, what specific change would fix it. "Make it better" goes nowhere. "The opening is too abstract; start with the specific example from paragraph three" gets you exactly what you need. This is the taste muscle. It strengthens every time you use it. And when you notice yourself giving the same feedback repeatedly? Write it down in your Skill's Gotchas section. The back-and-forth becomes the Skill (with a capital S).
The best AI skill is knowing yourself well enough to collaborate with a tool that amplifies whatever you bring to it. Bring clear thinking, get clear output. The AI is a mirror. Metacognition is learning to see what it reflects.
