Introducing Jev Skill Suggestion for Claude Code Now you can keep your skills completely out of the context window until they’re actually needed Set your skills as user invocable only, and the Mod sends the available skill list to Jev For every request, Jev classifies which skill best matches... the task, then injects only that skill into Claude’s context Works with both the TypeSafe AI API and Vercel AI Gateway Install it with one command: npx claude-code-templates@latest --mod productivity/jev-skill-suggestion Full implementation details on the Mod page atshow more

user says "make staging match prod." the skill is named environment promotion. curious how often the right tool stays invisible because the description and the request share no vocabulary.

@dani_avila7 That's a game changer for managing context. I remember struggling with this when trying to streamline processes. Makes customization way smoother!

how many skills can jev realistically hold before the classification step itself becomes the bottleneck

Cool pattern. I’d want it to be able to say ‘none’ though. A confidently wrong skill is probably worse than no skill loaded.

按需加载技能，终于不用把上下文塞满了

Interesting idea 🙂 Something I've wanted to ask for a while: how do you keep track of what's installed from aitmpl and where (~/.claude vs project .claude/)? And how do I know when a component I use got updated and is worth updating? Re-running @latest just overwrites files.

So we save context window space by spending an extra request deciding what goes in it. Balanced.

把 skills 按需拉进来 上下文窗口马上干净很多 工具越多越不能全塞进去 下一步就是让 agent 自己会挑对的 skill

Smart approach, but the failure mode worries me: if the classifier picks the wrong skill, or none when one was needed, the agent quietly loses context it would have had. Is there a fallback, and how do you measure classification accuracy?

keeping the unused skills out of context is nice. does it pass the list to the model on every request or cache it?

This is the same trick as lazy loading MCP tool schemas, keeping tokens out of the prompt until a classifier decides it needs them. How much latency does that classification step add before Jev picks a skill?

That is great but @AnthropicAI @trq212 have to solve the very long answer issue. I even give a clear prompt to get a short answer it gives a long one, as in this video example

does it actually inject the skill, or just suggest one?

the link is very helpful, I love to search

Context bloat has been the quiet tax on every agentic system. Selective loading based on actual invocation is the only way this scales past 50 skills.

skills stay out until Jev picks one. damn

Each per-request classification can be correct while the cumulative path still drifts. A skill absent at step one quietly changes how the agent reasons by step four, no visible miss.

Did you log how many tokens the skill list itself costs per turn versus injecting one skill, and how often Jev picked the wrong skill on a real repo?

By default only each skill's short description sits in context, and the model can still pick one halfway through a task, once it's read the code. A classifier on the prompt has to choose before any of that happens.

I dont know about this approach. Whats the chance of jev finding appropriate skills at the right timing vs SOTA models? Would you care to do some bench?

The useful pattern is lazy-loading for agent skills: keep the catalog cheap, classify the task first, then inject only the relevant capability. That reduces context noise without hiding what the system can do.

keeping system prompt overhead low is the real bottleneck here

Keeping tool schemas out of the system prompt until needed is such a clean way to dodge context rot on long runs.