What changed
A new AI for the small calls inside your workflows
On September 15th a company called TypeSafe AI released Jev, a new kind of AI model built to make quick decisions.
ChatGPT and Claude are large language models, or LLMs. You ask them something and they write you an answer. Jev does a narrower job. You give it some information and a question with a fixed set of possible answers, and it hands back one of those answers along with how sure it is. The question can be a choice ("is this email urgent, routine or junk?"), a score ("how good a fit is this lead, from 1 to 5?") or a yes or no ("do these notes say the contract is signed?").
If you've built skills or workflows with Claude or ChatGPT, they're full of small calls like these: sort the inbox, match a meeting to the right client, check a contract, grade a list. An LLM can make each call, but it's slow and expensive for the job, and it doesn't give you a confidence number your workflow can use. TypeSafe says Jev answers in under a second for a small fraction of the cost. Those are TypeSafe's own claims. For now it's invite-only, and a developer connects it to your workflows.
We got in and wired it into 6 of our own workflows. Two results stood out. A check that reads our client notes had been counting a contract as signed when the note said it was "to be prepared". Once Jev was reading those notes, it flagged that one straight away. Another workflow matches meeting notes to clients, and Jev caught a call whose title used a shortened client name our old matching had missed.
It has limits. We wrote 10 made-up emails based on the rules we use to sort our own inbox, and Jev got 8 right on its own. The 2 it missed were special cases: a "your email bounced" notice, and a group email where we'd only been copied in. Rules like those exist because something went wrong once, and Jev had no way of knowing them. It also sounded just as sure about the 2 it got wrong as about the rest. John Berryman, an AI consultant who reviewed Jev independently, found kinds of questions where Jev's confidence numbers don't hold up. Our written rules cover exactly the 2 cases it missed, so they stay as fixed instructions Jev can't overrule. On our real inbox it's only watching for now, and a person still makes the call.
If you're looking to improve the quality of a lot of the output of your skills and workflows, then Jev is a good way to hand off a lot of the low-level sorting that is expensive on an LLM and that Jev is specifically designed for.
Sources: TypeSafe AI, "Introducing System One Models & Jev", September 15th · John Berryman, Arcturus Labs, "Will OpenAI Eat Jev's Lunch?", September 21st
Skills you save in the Claude app now reach Claude Code
Claude lets you save skills: reusable instructions for a job you do often, like your weekly report or the way you write proposals. Claude Code, the version of Claude built for writing software, now picks up any skill switched on in the Claude app, as long as it's the same Claude account. Set it up once and it works in both places.
Try: switch on one skill in the Claude app, then open Claude Code and type /skills. The skill you switched on should be listed there.
One instructions file now works across your AI coding tools
Most AI coding tools read a file in each project that says how your team works: which folders matter, how you name things, what never to touch. Claude Code reads a file called CLAUDE.md. The others mostly read one called AGENTS.md. Claude Code now reads AGENTS.md when a project has no CLAUDE.md, so a team using more than one tool can keep a single copy of its rules for all of them.
Try: if your developers use more than one coding tool, ask whether they keep 2 instruction files. If they do, they can merge them into AGENTS.md and drop CLAUDE.md.