Claire reports convincing creative results with Fable 5.2, but still refuses to delegate a website to it alone. At ChatPRD, she shut down free trials after costly abuse. Mathematics educator Scott Farrar chose a different form of control for his home: searchable files rather than memory locked in conversations.
Claire reserves Fable for complex or creative missions
A complex animation completed in 16 minutes, exercises often completed in less than a minute: Claire reports these times after trying Claude Fable 5.2, from Anthropic. She also appreciates her limited interventions at moments when human action becomes necessary. Her judgment contrasts with that which she had on the first Fable, difficult to understand and not very pleasant to use, then on Opus 5, which had not convinced her further.
For a marketing video, she provided him with a tweet as well as graphic instructions from CXO.dev. According to her, the result tells a story with personalized invader characters, capable of walking, jumping or clapping, where other models tried had delivered nothing comparable. She now uses this exercise as a reference test. On another level, vague photos of a dismantled piece of furniture, without instructions, would have been enough for Fable to recognize the elements, draw a diagram and propose an assembly followed successfully by his fireplace, after the failure of other models.
Claire also entrusted him with the resumption of a complex ChatPRD pipeline and reported the creation of a functional 3D typewriter. She appreciates the humor of a game devoted to bugs and engineering trade-offs, extended by a trailer where the finance department imposes a cheaper model after discovering the token spending. She considers this version more accessible, quick and creative, without delegating the independent design of a website to her. His preference remains Opus 5.5 for most of the code and knowledge tasks; she would choose Fable for larger, difficult or creative missions.
ChatPRD closes its promotional access after abuse
At ChatPRD, Claire estimates the damage from token fraud at tens of thousands of dollars, plus a large portion of her engineering time. A technically experienced solo founder, she nevertheless manages a SOC 2 compliant product that is regularly subject to penetration testing. Each fraudulent use of a model represents an effective expense for the supplier.
It distinguishes six categories of abuse, including backdoor trials, promotion resale, card testing, product relaying and MCP exploitation. According to his account, automated account creations provided access to the most expensive models to perform work unrelated to ChatPRD. Neither the credit card requirement nor paid trials were enough: it eliminated free trials, forgoing a channel it deemed valuable for product adoption and enterprise leads.
Partner promotions were also stolen and resold to fuel other AI companies, Claire claims. Estimating that about half of its codes had reached resellers, it stopped issuing them. It now recommends continuous monitoring of spending, rather than just counting requests, closing free entries and verifying the customer beyond their account. In his experience, abuse quickly stops as soon as expensive models become inaccessible to fraudsters.
The evening letterFree · 7 p.m.
Do you want the best AI tools before the rest?
We test and decipher the new AI tools every evening, in 5 minutes. Free.
Included upon registration: our selection of the best AI guides & comparisons.
Choose your pace
Free · No spam · Unsubscribe in 1 click
ChatPRD classifies suspicious sessions with Jev
For sessions crossing its abuse threshold, ChatPRD uses a non-data curation Jev loop, with low-cost decision models that classify activity in real time. Confirmed fraud results in a ban or a discreet switch to a cheaper model. According to Claire, a user had also transformed the personalized MCP support into a computing resource available 24/7; it limited admitted hosts and blocked free tunnels as well as infrastructure without apparent connection to a legitimate company.
After massive exclusions, clumsy appeals rubbed shoulders with courteous messages produced by AI, she says. Its support agent now has anti-fraud tools and guidelines explicitly prohibiting the reactivation of accounts or the awarding of promotions. Claire also reports a sharp reduction in abuse after the blocking of certain countries with Vercel Firewall, a decision which led it to re-examine its international strategy. She now believes that trust is earned rather than assumed.
Claire claims to have identified the same fraudsters behind rotating proxies across several major AI services. She began to speak publicly about these practices. With Codex, she also prepared cybercrime reports filed with the FBI.
Scott Farrar keeps context in auditable files
Math educator and parent Scott Farrar put together Claude Code and Obsidian to manage a shared memory in the home. He prefers the context preserved in Markdown to long conversations, of which he criticizes the memory drift: the files remain readable, correctable and reusable from one session to another. He provides raw information and lets Claude tidy it up, while a custom-made linter prevents agents from making the documents too long, especially the chore one.
Asked about housing documentation, Claude suggested he photograph the devices. These images fed into a searchable inventory including their references, serial numbers and warranties. Scott Farrar sees it as a conversational version of his father’s filing cabinet, which kept receipts and warranties and knew how to explain repairs made decades before. When the agent deviates from the task, it starts again or inspects its reasoning before erroneous data enters the memory files.
The household shares its tasks without imposing everything on GitHub
Scott Farrar and his wife first wrote down the household chores that came to mind. Claude then divided them according to their priority, their cost and the necessary information; according to Scott Farrar, the device highlights delays before they turn into emergencies. Family projects and chores are tracked in a common list, with a wall-mounted screen showing planned trips, remaining tasks, recent accomplishments and suggestions for the day.
Sharing did not entirely follow Claude’s recommendation, which suggested GitHub to prevent file conflicts. His wife finding this organization too complex, the couple reserves GitHub for common scripts and uses Dropbox for common documents. For insurance, Claude also compares the quotes with the characteristics of the accommodation, points out possible gaps in coverage and prepares targeted questions. Scott Farrar thus believes that he is less dependent on the sole explanations of the person who sold him the contract.
Scott Farrar wants tutors who require reasoning
In teaching, Scott Farrar favors assistants who ask students to explain their reasoning rather than solving the exercises for them. He wants them to remain encouraging while maintaining a high standard, without erasing the difficulties that he considers useful for learning. He also thinks that some children can more easily submit unfinished ideas to an AI than to a human interlocutor.

