Tech news picked on 2026-11-10
Claude Fable 5.2 prompt guide: reducing early stopping and unnecessary approval requests
- Anthropic's official documentation lays out how to write instructions suited to Claude Fable 5.2. It says the effort level, a setting that determines how much the model thinks, should start at the default of high and be raised or lowered based on your own evaluation. It advises re-comparing values tuned for earlier models rather than reusing them as they are.
- In long tasks, the model sometimes stops with work left over, and this happens more often at low and medium. When Anthropic re-ran recorded sessions at medium, adding a phrase telling the model to finish the job cut sessions that ended with work remaining by about half, and output tokens increased by about 30%.
- When the prompt stated that the code was the model's own, the rate of merging duplicate code rose to about half of attempts, and without that statement it was below 1 in 10. This is the result of a small coding test by Anthropic, and in it output tokens rose by about 1.5 times and unrequested changes also appeared.
- Approval rules are applied exactly as written, so the target actions must be specified precisely. In an Anthropic simulated test, when sending messages to outside parties was listed as requiring approval, the model did not carry out about half of routine requests to send messages to colleagues or other teams.
- This is a chatbot that finds evidence in the bylaws, notices, and contact messages to answer questions such as those about fees or refunds. It splits documents into small pieces and shows the AI only the pieces closest to the question (RAG). The experiment documents are materials for a fictional club created by the author.
- When the bylaws were split by article instead of cutting documents every 400 characters, the accuracy rate on 32 questions rose from an average of 52% to 72% over 3 runs. The author disclosed that grading was done by the same Claude Sonnet 4.5 that wrote the answers, so it may be lenient toward its own answers.
- Running keyword matching search alongside, guaranteeing at least 2 bylaw pieces, and raising the number of pieces passed on to 8 were all within the margin of error or made results worse. Because principles and cases are written with the same words, pulling up the bylaws pushed out the needed contact messages.
- When the author switched to a local model for a laptop (RTX 3050) to avoid sending data externally, the accuracy rate dropped by about 2 tenths. Replacing the search embedding reduced results by an average of 2.7 questions, and replacing the answer model with a 4 billion parameter model reduced them by about 13 more questions.
- Even if all tests pass, you cannot tell whether those tests actually catch mistakes. Mutation testing deliberately breaks the code in small ways and then checks whether the tests fail. If they fail, the mutant was caught; if they pass, it was missed. Stryker is a tool that automates this process.
- The author's team applied it to a Prisma client (131 models) that automatically filters data per company. The definition table listing per-model conditions was 100% from the start, and the hand-written client tests became 100% after being supplemented at 89.74%.
- Among the missed cases was a test where an attempt to delete another company's row failed because of a foreign key constraint, not the per-company filter. The bind call the AI added following a review comment turned out to be unnecessary code, since removing it left behavior unchanged.
- In the author's environment, 407 mutants took 3 minutes 30 seconds, and when the scale grew to 1,326, it took 24 minutes. They had to specify only the real-DB tests so unrelated tests would not run, and the author is considering running it once before a PR.
- An engineering manager (EM) is a manager who leads a development team. The author says many developers probably see managers only in terms of approvals, meetings, coordination, responsibility and pressure, and admits that they themselves once turned down an offer to become a manager.
- The author argues that a manager is not a role about managing people only. It is a role of solving organizational problems so that the team can perform well, and the author sees the process of finding problems, understanding the structure and pinpointing the real cause as similar to engineering.
- As an example, the author describes investigating why team A often had to coordinate with team B. They found that the code owned by the two teams was tightly entangled, and after it was loosened, team A could develop without coordination.
- The author recommends that an EM of a single team act as a playing manager who also develops directly. The reason is to build technical judgment, but the author also admits that doing both is a heavy burden and that development skills grow more slowly.
Instagram unveils an effect builder that creates AI filters for Stories from a text description
- Describe the filter you want in text, and an AI chatbot builds a custom filter for Instagram Stories. You can also attach a style image for reference.
- You can upload your own photo to preview the result and revise it with follow-up requests. There is also an option where Instagram's AI rewrites your prompt in more detail.
- When a reporter at The Verge tested a 'filter that turns a photo into a comic book page', the result from the prompt the AI expanded into several paragraphs reportedly looked almost the same as the result from a one-line prompt.
- You can submit the filter you made directly from the tool. According to Instagram, effects that pass a safety review can be used by anyone in Stories, and each use links back to the creator's profile.
- Payments Canada, Canada's payment operator, has launched RTR, a new nationwide payment network. Participating banks and payment firms exchange payments and settle them immediately, 24 hours a day, all year round, regardless of business hours.
- It uses existing banks and payment channels as they are. According to Payments Canada, businesses can see cash flow in real time and attach more information to payments.
- Fraud prevention services are mandatory from launch. They include payee confirmation (offered only for consumer accounts on day one), fraud risk scoring, a risk list, and a fraud reporting platform.
- The first participants number 8, including Wealthsimple, Koho and Bank of America. Major banks and others will join in stages, with a plan to make it usable by at least 9 of every 10 Canadians within 1 years.