Anthropic published Agent Skills as an open standard this fortnight. Microsoft, OpenAI, Atlassian, Figma, Cursor, and GitHub all agreed to use the same file format. The practical effect: the skill you write once to capture your voice or your reasoning now runs on Claude, ChatGPT, Cursor, Copilot, and the others. The AI tool you pick stops being a lock-in.

Anthropic released Claude Opus 4.7 on 16 April. The headline features got the coverage. The quieter line is doing more work: Claude is now measurably better at writing notes to files and reading them back later. Anthropic also removed several controls that used to live inside the model. Direction of travel: fewer knobs on the AI, more weight on the files around it.

Laura Entis for Every this week: 'As you climb the frame hierarchy, your role is less about communicating the mechanics of a problem and more about defining what the most important problem even is.' She is writing for developers. The argument generalises. Most professionals have quietly been promoted to managing AI, not using it. The job is briefs, reviews, holding standards. The promotion came with no training manual.

Researchers benchmarked 34,000 AI skills running in real conditions on Claude, Cursor, and Codex. Two numbers to sit with. Pass rates dropped from 55% to 38% once the AI had to pick the right skill from a large library. Even when the right skill was available, the AI loaded it only 49% of the time. Reliability is a selection problem. More skills is not the answer.

Two production AI teams published post-mortems this fortnight. Manus, on running agents at scale, and Microsoft's Azure site reliability team. Both arrived at the same sentence, almost verbatim: 'we realised we were building a context engineering system that happens to do X.' The pattern is now named. The practitioners calling it this are not selling it to you. They are reporting what actually produced the result.