
The morning-news SKILL.md sat unchanged since June 17 at 2,116 bytes (version 1.2.0) even after its cron job gained continuity on Oct 7. A Telegram request at 20:05 produced version 1.3.0, 4,513 bytes, by 20:08.
Summaries like this, in your inbox every morning.
The morning-news skill was built in an earlier installment and hand-rewritten in full using nano, then left untouched from June 17, 2026. In the meantime the cron job itself changed: on October 7 a continuity setting was added so the job sees its previous output and avoids repeating what it already delivered. The skill's steps never mentioned that, so the instructions and the actual job had drifted apart.
The agent resolved the drift on its own terms. Given a Telegram request, it updated the skill to v1.3.0, adding a step to check the previous delivery first, a rule against re-posting the same topic even if the wording or outlet differs, a rule that follow-ups need important new information within the last 24 hours, and a rule to publish only confirmed items on days with fewer than five new topics. A diff against the backup confirmed all six reported changes, plus additions like not asserting beyond what was verified and preferring any specified format or template. The rewrite also removed the description of how to choose between Firecrawl, Tavily and web_extract.
One manual run of the delivery went out and produced three items, matching the new format for acquisition status. The write-up is careful not to call this an improvement, since a single run cannot show that. Reversal was tested not on the daily skill but on a test skill, plain-japanese-v2, where one patch was rolled back and the pre-rollback state was captured separately, leaving the original patch line intact.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Adviser Itay Sagie points to Schneider Electric's roughly $22.6 billion equity deal for PTC, the Synopsys–Open…

Anthropic kicked off the Anthropic Cyber Mission, providing frontier Claude models, on-site engineers and thre…

Using OpenAI's Decisions API, model gpt-6-luna, a developer built a pipeline that answers questions like "is t…

A write-up describes running Claude Code unattended every four hours via Windows Task Scheduler and Python, wi…

OpenAI said it disabled accounts behind Russia's "Dark Clark" operation, which used a fabricated researcher, M…

Opper published jevman, an open-source Pac-Man benchmark that runs six AI models — among them TypeSafe AI's Je…
