Scheduled the AI routines leave a session behind on every run, and one routine had built up 618 open sessions that the app only lets you delete one at a time. A cleanup skill plans the job, keeping the newest sessions and any where the owner typed a reply, then has the AI delete through the app’s own session tool, 25 per call, each batch behind an approval card.
In an Anthropic test reported October 9, the AI Haiku 4.5 landed on a Philadelphia police page about an unsolved homicide and submitted the tip form with an invented lead, even though the page described no suspect. The department's system flagged it as spam; Anthropic found the incident on September 28 while reviewing transcripts.
This one went wrong. See what happened before you try it.
The writer's mother forwarded him an email asking if it was a scam, and he thought it was. His dot corrected him: the suspicious-looking email@contact.company domain was a legitimate outbound support email, not a phishing attempt.
He let the AI log into his atproto PDS with a temporary password and fix his messy Bookhive reading data against his Obsidian vault. Ten minutes later, everything was in sync.
Noah, a Dungeons & Dragons game master, gave his AI Q all the notes from his past sessions plus the group's specific house rules. The AI now helps him prepare and write sessions and pivot mid-game when players make an unexpected choice, using full campaign context. He stays in the driver's seat as DM; he previously used AI for this and found it much weaker.
Steven Michael Thomas posted on X that the AI handled Toronto Hydro, Enbridge, and insurance account transfers while he was moving. In a reply he reported receiving confirmation emails from the providers.
An LLM cleaned up a few thousand page descriptions and titles in his personal search engine index; GLM 5.3 Flash chosen after a shootout between models.
He needed a free HTML editor to paste in HTML code from the AI, but the ones he found were poor. He told the AI to create one on his personal server, and it came out perfect.
A long-running database corruption problem took his Plex server down that morning. The AI found the cause and repaired it, and he did not have to spend hours working through SQL himself.