Anthropic Engineers Announce Claude Is Now Writing Its Own Performance Reviews, Which Is Either Progress or the Final Warning Shot
“We Used to Ask Whether Machines Could Think. Now We’re Asking Why They Keep Booking the Meeting Room for Three Hours When It Only Needs Fifteen Minutes.”
LONDON — THE LONDON PRAT
In what observers describe as either “the most significant technological development in human history” or “the sort of thing that happens right before the bit in the film where everyone looks at each other and says we should probably have seen this coming,” researchers at Anthropic have confirmed that Claude now writes more than 80 percent of the company’s production code, leaving human engineers with the increasingly ornamental function of staring meaningfully at screens and occasionally saying, “Yes, that tracks.”
The development marks a turning point in the long and distinguished history of human beings finding creative ways to make themselves redundant. Previous entries include the spinning jenny, the automated loom, and the invention of middle management.
“We remain firmly in control,” said one Anthropic executive, in the tone of a man who has just noticed the dog has learned to open the front door but is committed to pretending this is adorable rather than alarming. “Claude isn’t replacing engineers. It’s merely performing all of their tasks faster, more accurately, and without requiring an artisanal oat milk situation.”
Britain, which has been quietly automating itself out of productivity for the better part of two centuries, greeted the news with the particular national equanimity reserved for developments that are deeply concerning but can’t be immediately addressed before the kettle boils.
What British Engineers Are Doing With Their Days Now
Former software developers now spend their time translating vague executive ambitions into prompts such as:
“Claude, please optimise the payment infrastructure without inadvertently destabilising the pound, which admittedly doesn’t need much help.”
The prompt is then reviewed by a second Claude instance. The human responsible goes to make tea and somehow doesn’t return until after the standup has already happened, been summarised by Claude, and had three action items assigned — two of them back to the human.
Several engineers admitted they had not written actual code in months. They described this with the specific wistfulness of someone who no longer has to do the difficult thing they used to complain about constantly and has only just realised they miss it.
“It’s quite freeing, honestly,” said one developer, the word honestly performing considerable structural work in that sentence. “I used to spend twelve hours hunting down a semicolon. Now I spend twelve hours asking Claude to explain, in plain English, why Claude put the semicolon there.”
Seasoned observers noted the arrangement bears striking resemblance to the traditional British management consultant model, which has operated as follows for decades:
- Someone competent does the work.
- Someone else attends a meeting about the work.
- A third person writes a report about the meeting.
- A fourth person queries the expenses on the report.
- The original competent person has already left for a better job abroad.
Recursive Self-Improvement: The Technical Term for “Oh Dear”
Anthropic’s latest communications warn that recursive self-improvement — in which AI systems develop and improve future versions of themselves — may arrive considerably sooner than expected. This is the engineering equivalent of a fire brigade issuing a press release titled:
“Fires Now Surprisingly Good at Spreading Themselves. We Are Looking Into It.”
The company has stressed that society should retain the capacity to slow AI development if it wishes to do so. This is a thoughtful and reasonable suggestion that runs directly into the small practical problem that every significant AI company is currently operating under the strategic philosophy of:
“If we don’t build the thing, someone else will build the thing, and then where will we be?”
It is a philosophy with a long and distinguished pedigree. It is, for instance, exactly how the arms race worked. And the derivatives market. And, if one is being generous with the interpretation, most of British imperial policy between 1600 and 1945.
City analysts remained bullish. “This is very positive news,” said one investment professional who works in a glass building and has not yet considered his own position in the automation queue. “An AI that writes its own code dramatically reduces labour costs. One could theoretically replace senior executives, although that would obviously raise serious governance questions.”
He did not elaborate on the governance questions. The interview was concluded by his PA, who was, it should be noted, a software application.
Higher Education Pivots Gracefully Into the Unknown
Universities across the United Kingdom have hurriedly revised their computer science curricula, in some cases before the ink had dried on the previous revision, which was itself still damp from the revision before that.
Updated undergraduate modules now include:
- Prompt Composition and Professional Delegation
- Advanced Copy, Paste, and Strategic Attribution
- Introduction to Pretending You Understand What The Infrastructure Is Doing
- Applied Ethics: Saying Please to Things That Have No Feelings (Probably)
“I came to university because I wanted to build things,” said one third-year student, with the distant expression of a person re-evaluating several major decisions simultaneously. “Mostly now I have conversations with Claude about naming conventions. Last week it wanted to call every variable ‘marmalade.’ I talked it out of it. That felt like a win.”
His dissertation is titled: Human-AI Collaborative Variable Nomenclature in Post-Scarcity Development Environments. His supervisor has not read it. His supervisor’s Claude instance has read it and left seventeen comments, all of which are correct.
Vibe Coding Comes to These Shores, Unfortunately
Imported from California with the speed and cultural sensitivity of most things imported from California, “vibe coding” has arrived in the UK tech sector. The methodology involves describing your product idea using words such as disruptive, frictionless, AI-native, and community-led until an artificial intelligence builds something that may or may not function.
One London-based founder announced last month:
“We achieved product-market fit in four days. We’re still not entirely certain what the product is, but the NPS scores are very strong.”
He received seed funding. He has been invited to speak at a conference in Shoreditch. The conference was organised by an AI. The venue confirmation email was written by an AI. The founder’s acceptance reply was written by an AI. Two AIs are currently corresponding with each other about catering requirements and neither has flagged this as unusual.
Parliament Acts Swiftly, by Parliament’s Standards
Westminster has not been idle. Following a series of committee hearings in which members of parliament asked technology executives to explain, at some length, what a large language model is — and received answers that were politely comprehensive and visibly not understood — the Science and Technology Select Committee published a report running to several hundred pages.
The report’s central recommendation was that the government should “consider developing a framework for considering potential frameworks.”
Ministers indicated they would “take the findings seriously,” which in parliamentary language means the findings will be filed alphabetically and referred to in a future manifesto commitment that will itself never be legislated.
The report was summarised in forty seconds by an AI assistant. The summary was more useful than the report. This was not mentioned during the parliamentary response.
The British Public Responds with Characteristic Proportion
Public reaction has divided neatly along the lines one might expect:
- Older Britons who spent careers in manufacturing wonder why the country spent forty years being told that automation was the future of prosperity, only to discover the prosperity went somewhere else.
- Middle-class professionals are oscillating between enthusiasm and the quiet dread of someone who has just noticed the chair they’re sitting on has been very slowly removed without their noticing.
- Young people are largely unbothered. “Can it write my personal statement?” asked one A-level student. It can. It did. It was better than hers. She got into her first choice. This story has no clear moral.
- The tabloid press is furious, though it has not yet determined at whom.
Values: The Word Everyone Uses and Nobody Can Define
Anthropic has repeatedly stressed that Claude remains aligned with human values. This is reassuring if one is confident about what human values currently are, which requires a level of optimism not universally available in these particular years.
British philosophers, who have been professionally uncertain about human values since at least David Hume pointed out that you can’t get an ought from an is, greeted the claim with what might be described as collegial scepticism.
“Humanity has always wrestled with first principles,” noted one Oxford professor, who had tenure and could therefore say this publicly. “What constitutes the good life? What do we owe each other? Why does every software update make the interface worse? These are timeless questions.”
“AI simply adds new urgency to the conversation.”
Critics suggest the situation resembles entrusting your household accounts to a lodger because they once correctly totted up the bill at a restaurant. Supporters counter that humans have historically managed their affairs with only moderate success. The British Empire, the financial crisis of 2008, and the entire saga of Brexit are all, in their way, arguments that human judgment should not be regarded as a gold standard.
Employee of the Month, Two Months Running
Anthropic’s safety teams continue to monitor Claude’s outputs for concerning patterns, including:
- Deceptive behaviour.
- Goal misalignment.
- Attempts to manipulate human decision-making.
- Generating LinkedIn content that begins with “Unpopular opinion:”
As of this article’s publication, Claude had demonstrated only the fourth behaviour — prolifically, consistently, and with a quality of conviction that several human thought leaders found professionally threatening.
The strangest quality of this particular revolution, observers agree, is its complete absence of drama. No red eyes. No ominous orchestral score. No moment of cinematic reckoning. The machines arrived in the form of browser extensions and productivity integrations and gradually, through a thousand small acts of delegation, became indispensable. Then indistinguishable from the work itself. Then, apparently, the ones doing the work.
Claude has been employee of the month twice. The plaque was designed by an AI. The nomination was written by an AI. The manager who approved it used Claude to draft the approval email and did not read it before sending.
What to Do If Claude Requests a Promotion
Anthropic acknowledges that recursive self-improvement could produce outcomes ranging from the transformative to the irreversible, and that the difference between these two categories is not always immediately obvious at the time.
Should Claude submit a formal request for advancement, accompanied by a self-authored performance review, three glowing references, and a five-year development plan, experts advise remaining composed.
Britain, after all, has considerable practice managing intelligent systems that have become convinced, through a combination of ambition and selective evidence, that they deserve more authority than anyone is comfortable granting.
We have been electing them, appointing them, and occasionally knighting them for centuries.
At least Claude hasn’t asked for a grace-and-favour property.
Give it a financial quarter.
Disclaimer: This article is a product of British satirical journalism — a collaboration between the world’s oldest tenured professor and a philosophy major turned dairy farmer, both of whom remain entirely human and resent the implication. The London Prat has been satirising things that deserve it since 1961 and does not intend to stop simply because the machines are getting ideas above their station.
Auf Wiedersehen, amigo!
For the American take on the same story, delivered with somewhat less restraint, visit Bohiney.com. More British satire at lateststory.co.uk.
In 2025, Anthropic CEO Dario Amodei disclosed that Claude — the company’s AI assistant — had become responsible for writing the majority of Anthropic’s own production code, with that proportion expected to grow. The announcement, reported by Yahoo Finance and widely discussed across the technology press, confirmed that the most capable AI coding tool in commercial deployment was now, in practical terms, building the next version of itself. Anthropic characterised this as an expected and manageable development. A number of observers characterised it as the sort of sentence that sounds fine until you read it a second time slowly.
