A few small changes in Claude Code this week that are genuinely worth knowing about, plus one story that reaches well beyond the tool itself.
/verify and /code-review stop invoking themselves
2.1.215 changed the behavior of two specific, bundled Anthropic skills: /verify and /code-review. Previously, Claude Code could decide on its own that a moment in the conversation called for one of them, and just run it. Now both are invoke-only: you have to call them manually. This isn’t a change to skills in general, it’s a deliberate call from Anthropic that these two particular, longer-running checks shouldn’t get to decide on their own when to spend your time and tokens.
Your own custom skills aren’t affected by this, and were never left to guesswork in the first place. Whether Claude can invoke a given skill on its own is controlled by two frontmatter fields in SKILL.md, independent of this update: disable-model-invocation: true means only you can run it, Claude never will on its own - the docs explicitly recommend this for skills with side effects, like committing, deploying, or sending a message, so Claude doesn’t decide to do it just because “the code looks ready.” The opposite is user-invocable: false - only Claude can use it, and it won’t show up in your / menu, which fits skills that are pure background knowledge rather than an action. If you’re building skills that actually change something (writing files, committing, sending messages) and haven’t set disable-model-invocation yet, this is a good moment to do it, regardless of this specific news item.
More control over filesystem isolation
A day later, in 2.1.216, independent control over filesystem isolation landed. In practice that means more flexibility in exactly how you run your code and what Claude Code can touch while it works, instead of one all-or-nothing setting.
That matters most when you’re juggling several projects at once, or running subagents in isolated workflows where you want a guarantee that one process can’t reach into files that belong to another. Before this, you had to trade off convenience against safety. Now you can match the isolation level to the situation you’re actually in. If you handle several client projects out of one working directory, that’s a concrete reason to review this setting now, before one careless agent run touches something it shouldn’t.
One small nice-to-have
2.1.217 added emoji shortcode autocomplete to the prompt input, along with better warnings when you type an invalid one. Nothing that changes your workflow, but nice to see someone at Anthropic still sweats details like this, even on a tool that spends most of its day writing code rather than chatting in emoji. Actually useful when you’re writing docs or a commit message and want to drop in a specific emoji from memory instead of copying it from somewhere else.
To end on: a shorter road to a rare disease diagnosis
Anthropic announced grants for rare genetic disease research: up to $50,000 in Claude credits over six months, for research teams. Anthropic put the reasoning plainly - AI makes it possible to accurately model rare genetic diseases and detect patterns across them that a single research team might simply miss.
That matters because roughly 300 million people worldwide live with a rare disease, and diagnosis alone often takes five years or more, regardless of whether you happen to live near a good university hospital or not. A model that can compare a patient’s symptoms against thousands of similar rare cases at once could genuinely shorten that wait, especially in places with no local specialist for that particular condition. Hard not to be glad someone’s pointing compute in this direction. If you know anyone at a university or research institute working on a rare disease, this one’s worth forwarding: $50,000 in credits over six months of research is a concrete, usable amount, not just a PR line.
