The Job, Repriced

In January 2023, Anthropic posted a job for a “prompt engineer & librarian” at $175,000–$335,000 a year, a listing Fortune covered under the headline of an engineering job where you do not write code. In July 2026, Boris Cherny — creator and head of Claude Code — told an interviewer that prompt engineering was “not that important,” and Search Engine Journal relayed the comment as news. Between those two dates a job title was born, priced at a sixth of a million dollars, and retired.

Around the same stretch of 2026: a research organization found it could no longer cleanly run a study that required developers to work without AI, because 30–50% of them declined to submit such tasks. Klarna, which in 2024 said its AI assistant did the work of 700 full-time customer-service agents, spent 2025 rehiring humans for the same queues. JD.com reported more than 70,000 sellers running livestreams with digital humans that never sleep. And Chinese universities, in the same Ministry of Education overhaul that revoked 12,200 degree programs, added AI programs by the thousand — including, per the Chinese columns, programs for governing AI content.

Our Chinese-language columns ran these as separate dispatches — the death of a job title, the four nets AI weaves around wallets, jobs, eyes and confessions, the streamer economy meeting the governance major. Read together they are not three stories but one machine, and this piece takes it apart as such: how work is being repriced — the price of producing something falling while the price of vouching for it rises.

The gears

The mechanism deserves a name: the verification premium. When a system can generate a draft, check it against a stated criterion and revise itself, two prices move at once. The premium on getting instructions right the first time collapses — the system will iterate. The premium on judging whether the output is any good rises — someone still has to stand behind it when it ships, streams or bills.

Three gears, then. First, instruction deflation: the craft of precisely wording a request was worth $335,000 in 2023 because a model that misunderstood you would run the error to completion; an agent that checks itself makes first-shot precision a convenience, not a moat. Second, standard-supply deflation: when generation gets cheap, routine output — boilerplate code, product-pitch livestreams, formulaic copy — floods in at near-zero marginal cost, and the human wage attached to producing it compresses. Third, verification inflation: every cheap draft needs a check, every synthetic face needs a label, every AI-assisted customer needs a human escalation path — and the jobs, budgets and university seats attached to checking expand.

The Chinese column on the “four nets” compressed the same movement into one sentence about the software trade: work is shifting from writing to reviewing, from creating to gatekeeping — the divide running not between AI users and AI refusers, but between the person who generates and the person who signs off.

Instance one: the two-year job

The working side of the prompt-engineer story is a clean arc with receipts. The job existed: Anthropic’s own listing in early 2023, salary band $175,000–$335,000, asking for a “hacker spirit” and no traditional coding. The job’s premise was real: early chat models lost context between sessions and ran first errors to completion, so the cost of a mis-specified request — re-explaining, re-aligning, re-correcting — landed on the requester, and a specialist who could compress that loop into one good prompt captured part of the savings. The job’s ending is dated and on the record: in a July 2026 interview relayed by Search Engine Journal, Claude Code’s head said prompt engineering was “not that important,” and described giving the model a task, a way to verify its own output, and an instruction not to stop until done — one line where 200 words of constraints used to go. The column’s example, relayed from the same interview: a request to rewrite an Electron app as native Swift, with the agent told to screenshot and pixel-compare its own work until it matched [anecdote as relayed by our column].

The failing side, in the same period: the compensation the job existed to provide did not disappear — it moved. METR’s measurements (next instance) put the correction work back on the developer’s own clock, just wearing a different name. And the replacement skill is not “no skill”: specifying a task and its acceptance criterion — what counts as done, what counts as broken — is the new craft, which is why the same interview that retired the title immediately named a successor activity. One bookkeeping note on the column’s own narrative: its headline put the arc at twelve years of AI learning to work; its own timeline runs from the 2012 ImageNet result to 2026’s agents, which is fourteen.

Instance two: faster on the inside, slower on the clock

The working side of the productivity story is documented at the volume end of work. Klarna’s February 2024 announcement: its OpenAI-built assistant was doing the equivalent of 700 full-time agents, resolving issues in under two minutes against eleven for humans, with a 25% drop in repeat inquiries. The routine tier of service genuinely absorbed. At the hiring end, Indeed Hiring Lab measured job postings for junior tech titles down 34% — a decline that began in 2023, the year the assistant class shipped. And the review work is quantified: CodeRabbit’s analysis of 470 pull requests found AI-authored changes averaging 10.83 issues against 6.45 for human-authored ones — 10.83/6.45 = 1.68, roughly 1.7× — with security findings about 2.7× higher and logic errors 75% more frequent. Each line item is a task someone is paid to catch.

The failing side is the measured half of the same period. METR’s randomized trial (July 2025): experienced developers on their own mature repositories took 19% longer with early-2025 AI tools while believing themselves about 20% faster — they had expected a 24% speedup before starting. The follow-up, February 2026: 57 developers, and an “unreliable signal,” because 30–50% of them declined to submit tasks they would have had to complete without AI; one participant completed none. The no-AI control condition had become unenforceable — the dependency is no longer a preference. Klarna’s arc closed the loop: by May 2025 the CEO was telling Bloomberg the cost-cutting had gone too far, and the company was recruiting humans back into customer service for the queries where quality, not volume, priced the answer. The column’s supporting relays, kept tagged: an AI-company founder’s figure that 44% of AI spending went to fixing AI-generated bugs [unverified — single-company claim, no independent anchor found]; Uber exhausting its annual AI budget in four months of 2026 with no measurable output gain [unverified — internal figures as relayed via The Information].

Both columns, then, say the same thing from opposite sides: the generating got cheaper and the checking got bigger. The 34% and the 1.7× are the same economy measured at the job posting and the pull request.

Instance three: the anchor jobs and the gate jobs

The working side of the livestream story runs on arithmetic that survives translation: a digital human streams around the clock, in switchable languages, with no contract, no schedule and no persona to collapse — against a human anchor who sells by the hour and carries the wage of one. JD.com’s own program numbers, as relayed by industry tracking: 70,000+ sellers onboarded, costs cut to about a tenth of a human host, conversion up roughly 30% [platform figures]. A vendor case on Shopee Live in Malaysia reports a brand at 15× its previous broadcast hours — always-on against about twenty hours a month of human hosting — and +299% in live-sales revenue [vendor case-study figures]. The backdrop scale, per WIRED: more than a third of Chinese e-commerce already ran through livestreams in 2024. Our column’s domestic relays: GMV growth often above 200% year-on-year on major platforms, top digital sessions in the tens of millions of yuan, costs at one-fifth of a human host or less [unverified — platform-disclosure figures as relayed].

The failing side arrives as the flood’s invoice. Synthetic faces at scale mean synthetic fraud at scale, and the labeling-and-removal machinery is now publishing its own numbers: the Cyberspace Administration’s AI campaign, launched April 2026, removed more than 14,000 AI products and 5.61 million pieces of illegal AI-related information in four months, with a second phase announced in July targeting disinformation built with AI (China Daily, TechRepublic, Global Times). Our column cited the earlier 2025 round — 3,500 products, 960,000 items [as relayed]; the enforcement base rate has since quadrupled. The labeling measures for AI-generated content have been in force since September 2025. And the education system is repricing in the same direction: the Ministry of Education overhaul revoked or suspended 12,200 degree programs while adding over 10,000 AI-focused ones (Forbes), a change touching more than 30% of programs (SCMP). The column’s account of universities rushing to open “AI content governance” programs is kept as a relay — the specific naming is not anchored in English-language records — but the direction it describes is the documented one: seats moving from producing content to policing it. The critics quoted are Chinese: faculty redeployed from adjacent disciplines, graduates pointed at four different jobs — platform review, regulation, compliance, AI safety — under one major name, and AI programs already inflated from dozens of universities to hundreds [as relayed]. The column’s own caveat about the curriculum applies to its critics too: a governance program teaching this year’s rules graduates into a field where the 2025-round numbers were already obsolete by the 2026 round.

The ladder underneath

Two backdrop numbers frame how fast this is moving. Gartner sizes the enterprise AI coding-agent market at $10–11 billion in its 2026 review, with Anthropic, Cursor, GitHub and OpenAI as Magic Quadrant leaders — a category that did not exist when the prompt-engineer listing was posted. And in June 2026 Gartner published the cost side of the same curve: AI coding costs are predicted to surpass the average developer’s salary by 2028 as token consumption compounds. MIT Technology Review, cited by our column, locates the bottleneck in the same place the METR and CodeRabbit numbers do — no longer model capability but infrastructure, review capacity and the plumbing between them. The 14-year arc from the 2012 ImageNet result to 2026’s self-checking agents is short enough that one generation of workers has now been priced up, repriced down and retrained twice inside it. Our watch issue on the same force from the graduates’ side of the desk: the seniorization of entry-level hiring.

What outsiders usually get wrong

• That prompt engineering faded because people learned to write better prompts. The documented cause is structural: the target of the compensation moved. A checking loop inside the system replaced first-shot precision as the scarce skill — the interview that made headlines said so, and the METR numbers show the verification cost persisting, relocated to the developer’s clock.
• That Klarna’s AI eliminated 700 jobs. Klarna’s own two-step is the record: the February 2024 announcement claimed the workload of 700 full-time agents; by May 2025 the company was rehiring humans after quality fell. Jobs were re-scoped and repriced — the routine tier automated, the escalation tier re-staffed.
• That China’s AI-content rules exist only on paper. The labeling measures took effect 2025-09-01; the four-month enforcement numbers (14,000+ products, 5.61 million items) are published; a second phase was announced in July 2026. The argument inside China is about speed and staffing, not existence.
• That the dramatic percentages are settled fact. Several of the most-circulated figures — 44% of AI spend fixing AI’s own bugs, multiples-times-human conversion on Brazilian Shopee streams — trace to single-company claims and could not be independently anchored. The anchored numbers above (34%, 19%, 1.7×, $10–11B) are the load-bearing ones.

Sources

• Fortune: A.I. job listing calls for a ‘hacker spirit’ — Anthropic prompt engineer & librarian, $175K–$335K (2023-03)
• Search Engine Journal: head of Claude Code says prompt engineering not that important (2026-07)
• METR: measuring the impact of early-2025 AI on experienced developer productivity (2025-07-10)
• METR: uplift experiment update — selection effects, 30–50% withheld no-AI tasks (2026-02-24)
• Indeed Hiring Lab: experience requirements tightened; junior tech postings down 34% (2025-07-30)
• Bloomberg: Klarna turns from AI to real-person customer service (2025-05-08)
• Customer Experience Dive: Klarna reinvests in human customer-service talent
• CodeRabbit: state of AI vs human code generation — 10.83 vs 6.45 issues per PR
• The Register: AI-authored code needs more attention, contains worse bugs (2025-12-17)
• Gartner: enterprise AI coding agent market guide, 2026
• Gartner: AI coding costs to surpass average developer salary by 2028 (2026-06-24)
• WIRED: AI salespeople in China’s livestream e-commerce (2024)
• AnyMind: Bio-Essence on Shopee Live — +299% sales, 15× broadcast hours (vendor case)
• Forbes: China cuts 12,200 university programs, replaces many with AI degrees (2026-06-23)
• SCMP: China’s universities cut 12,000 ‘obsolete’ degrees amid AI race
• China Daily: 5.61 million pieces of illegal AI-related information removed in four months (2026-09-02)
• TechRepublic: China AI enforcement removes 5.6M content items, 14,000+ products
• CAC: Measures for Labeling AI-Generated Synthetic Content, effective 2025-09-01

Provenance & disclosure. This piece synthesizes three Chinese-language originals from our WeChat channel — “提示词工程师活不过两年：AI用12年学会的最后一件事叫“干活”” (2026-08-10), “被AI包围的2026：钱包、饭碗、眼睛和心里话” (2026-08-17) and “AI抢走了主播的饭碗，高校急着开“治理”专业” (2026-08-03; dates without prefix taken from the source-library file dates) — drafted with AI assistance under human editorial direction and adapted to English 2026-09-07. The “four nets” original is used here for its jobs net; its wallet, eyes and confessions nets are not load-bearing for this room. Verification: the Anthropic listing and salary band against Fortune (2023-03); the Cherny comment against Search Engine Journal (2026-07); METR’s 19%-longer / believed-20%-faster / expected-24%-faster figures and the February 2026 follow-up’s 57 developers, 30–50% task-withholding and one-zero-completion detail against METR’s own study page and update; junior tech postings −34% against Indeed Hiring Lab; Klarna’s 700-agent claim, sub-2-minute resolutions, 25% repeat-inquiry drop and 2025 rehiring against Bloomberg and Customer Experience Dive; CodeRabbit’s 10.83-vs-6.45 against its report and The Register; the $10–11B market and 2028 cost prediction against Gartner; the MOE overhaul’s 12,200 cuts and 10,000+ AI additions against Forbes and SCMP; the April 2026 campaign’s 14,000+ products and 5.61M items against China Daily and TechRepublic; the labeling measures’ effective date against the CAC’s official text. Unverified or relayed: the 44%-of-AI-spend-on-AI-bugs figure (single-company claim); Uber’s four-month AI-budget exhaustion (internal figures via The Information, as relayed); the Brazil Shopee conversion multiple; domestic GMV-growth-above-200% and one-fifth-cost figures (platform disclosures as relayed); the specific “AI content governance” major naming on the MOE list; AI programs’ expansion from dozens to hundreds of universities; the Electron-to-Swift anecdote (relayed from the same SEJ interview); the MIT Technology Review citation; the column’s 3,500-product / 960,000-item figures identified as the 2025 round of the same campaign. The Terminal-Bench scores in the first original were dropped: leaderboard ranks and scores diverge across harnesses (official board vs Vals AI vs Artificial Analysis) and the original’s specific 86.7/81.8/82.9 set could not be anchored. Ratios recomputed: 10.83/6.45 = 1.68 ≈ 1.7×; 2026−2012 = 14 years (the original’s headline said 12); Klarna’s 2-vs-11-minute resolution retained as announced. This is reported synthesis — not a SigPulse measurement, not career or investment advice. Our first-party measurements live in the dispatches and the /data/ ledger.