Eighteen months ago, putting "ChatGPT" on a CV did real work. It signalled you were early, curious, riding the wave before most of the workforce had opened the tool. Recruiters noticed. Interviewers asked about it. For a brief, strange window, familiarity with a chatbot was a differentiator.

That window has closed. Hiring data released in the last few weeks shows "ChatGPT" declining as a hired skill — down year over year and still falling month over month. One workforce analysis put it bluntly: listing it now sits in the same category as "proficient in Microsoft Word". Not wrong. Just baseline. The skill didn't become useless — it became universal, and universal skills don't get callbacks.

Meanwhile the demand underneath it moved — and it moved deeper. While surface-level tool familiarity stopped paying, postings asking for actual prompt-engineering capability quadrupled. The market didn't lose interest in the skill. It lost interest in the shallow version.

The speed of the collapse is the part worth sitting with. This isn't a skill that decayed over a decade, the way typing pools and switchboard work did. It's a skill that went from premium to baseline in under two years — fast enough that the people who added it to their CV during the hype are now watching it depreciate in the same document. If a two-year shelf life feels brutal, it's also the new normal for tool-adjacent skills: anything the tools make easier to learn will commoditise at the speed the tools spread.

Why the Self-Taught Path Hits a Ceiling

There's a structural reason most people's prompt skills plateau at the demo tier. Self-teaching follows the reward: you learn the techniques that produce immediately visible wins — the clever reframe, the persona trick, the magic phrase — because those are the ones that feel like progress. The techniques that produce *systematic* improvement — understanding why outputs vary, how instruction structure shapes reasoning, what makes a prompt robust across edge cases — are invisible in the reward loop. They don't feel like progress, so they don't get learned.

This is why the same people can use these tools daily for two years and still be surprised by their failures. Daily use at the demo tier doesn't compound; it just repeats. You get faster at the tricks you know. You don't get better at the craft underneath — because craft isn't absorbed through use, it's built through study. Millions of daily users plateaued at the same level for the same reason: use is not the same thing as learning.

The Skill That Commoditised Itself

What happened to "knows ChatGPT" is the oldest pattern in technology labour markets: a skill differentiates right up until the moment everyone has it. Millions of people spent the last two years self-teaching the same baseline — the same demos, the same clever-prompt lists, the same party tricks. Employers, for their part, trained almost nobody. The result is a market where the floor-level version of the skill is so common it carries no signal at all.

And the market repriced accordingly. At the same time the job postings asking for prompt engineering multiplied, the median salary attached to them dropped by roughly a fifth. That looks like a contradiction — more demand, less money — until you see what's actually being priced. The postings quadrupled because everyone now claims the skill. The salary fell because almost none of the claimants can demonstrate the version that matters. The market is flooded with the claim and starved of the capability.

The real demand, meanwhile, climbed the stack. Employers hiring in this space are no longer paying for people who can use the tools. They're paying for people who can build with them — the ones who wire AI into workflows, automate processes, construct agents. The skill that got you hired in 2024 was "familiar with". The skill that gets you hired now is "built with".

The split shows up everywhere once you look for it. The same hiring data that shows "ChatGPT" declining shows demand for the adjacent builder skills — AI workflow automation, AI SaaS development, agent construction — climbing fast. Same market, same quarter, two tiers moving in opposite directions. The tools didn't get less important; the version of skill that only touches the surface of them did. Everything above the surface repriced upward at exactly the rate the surface repriced down.

The Demo Version Everyone Learned

The reason so many capable people are stuck at the commodity tier is that the baseline version of prompting is genuinely easy to acquire — and feels complete. You learn a handful of techniques, watch the output improve dramatically from where it started, and reasonably conclude you've learned the skill. What you've actually learned is the demo: enough to make the tool impressive, nowhere near enough to make it reliable.

The professional tier is different in kind, not in degree. It's the difference between getting a good answer once and getting consistent, controllable, quality output across hundreds of variations. Between prompting as a parlour trick and prompting as an instrument — where you can steer tone, structure, reasoning depth and output format deliberately, diagnose why an output failed, and build repeatable patterns that work under real conditions. That depth is learnable. It just isn't learnable from the same listicle everyone else read.

And it's the only version employers are still paying a premium for. The generalist claim — "uses ChatGPT daily" — now reads as noise. The specialist claim — demonstrable, structured, deep fluency in making these tools produce professional-grade work — is what the postings that pay well are actually asking for.

"Show What You Built" Is the New Bar

There's a harder edge to the repricing, and it's changed what interviews look like. The hiring managers filling the well-paid roles in this space have learned — usually the expensive way — that claimed AI skills and demonstrated AI skills are different populations. So the bar moved: don't tell me you can prompt; show me what you built with it.

That's the filter most self-taught claimants fail. Not because they're incapable — because the self-taught path tends to produce a scattered, personal set of tricks rather than a systematic capability. You can produce good output; you can't necessarily explain why, reproduce it on demand, or adapt it to a brief you didn't choose. The gap between "I use it" and "I can show you a repeatable method" is exactly the gap the market is now pricing.

The demonstration bar is also why the CV line alone stopped working. When a skill can't be verified by a claim, the claim loses value — and the verification now happens live, in the interview or the take-home. The candidates clearing it aren't the ones who wrote the best bullet point about AI. They're the ones who can produce, under observation, output the rest of the room can't match. That's a different kind of preparation — and it's the kind a course teaches that a CV line can't fake.

It also explains why the employers paying the premium aren't the ones who trained their people. Most companies still aren't building this capability internally — they're buying it, which means the professionals who've built real depth are a thin pool being competed over. Thin pools set prices. That's the structural reason the depth tier holds its value while the surface tier commoditises.

And the premium isn't just salary — it's position. The people who can demonstrate systematic capability aren't competing for the roles that ask for tool familiarity; they're being pulled into the roles the tools created: workflow design, AI-augmented operations, the layer where the organisation decides *how* the tools get used rather than just *whether*. Those roles didn't exist at scale two years ago and they're exactly the ones the salary data shows expanding. The premium isn't paying for the skill — it's paying for the version of the skill that can build the next layer.

The Depth Tier Is the One That Compounds

The pattern that separates the repriced-up from the repriced-down isn't intelligence or technical background. It's whether the learning was systematic or incidental. The people still commanding a premium didn't just accumulate prompts — they built the underlying craft: how these models actually process instructions, why some framings work and others don't, how to engineer output that holds up under professional scrutiny.

That's what our Prompt Engineering for Professionals course teaches — a daily-email course in the depth tier: not another list of magic prompts, but the systematic capability the market repriced upward. For professionals ready to go deeper still — into the mechanics of how prompts actually steer the models — the Deep Dive in the Technical Science of Prompts takes the same craft to its foundations. And for the workflow-building tier the market is now paying most for, our AI Automation course covers the applied side.

The surface got free, so it stopped being valuable. The depth got expensive, so it's the tier worth owning. "Knows ChatGPT" was the version everyone could claim. The professionals still getting hired are the ones who built something the claimants can't show — and the gap between them is where the premium lives. It's still there. It's just deeper than the baseline everyone else learned.