Jev Skill Suggestion keeps your Claude Code skills out of context until one is needed

Claude Code sends the model a skill listing on every session: one line per installed skill, name and description. Each entry can run to 1,536 characters, and the whole listing is capped at 1% of the context window. If you have thirty skills installed and your prompt touches one of them, you are still paying for thirty. // Detect dark theme var iframe = document.getElementById('tweet-2101885477158547753-539'); if (document.body.className.includes('dark-theme')) { iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2101885477158547753&theme=dark" } jev-skill-suggestion removes that listing and puts a small decision model in front of every prompt instead. The model picks at most one skill, or none, and only that skill’s SKILL.md reaches the context window. What Jev is Jev is TypeSafe’s System One decision model. It does not write text. You give it a state and a set of typed questions, and it answers with calibrated probabilities in a few hundred

Jev Skill Suggestion keeps your Claude Code skills out of context until one is needed

Claude Code sends the model a skill listing on every session: one line per installed skill, name and description. Each entry can run to 1,536 characters, and the whole listing is capped at 1% of the context window. If you have thirty skills installed and your prompt touches one of them, you are still paying for thirty.

// Detect dark theme var iframe = document.getElementById('tweet-2101885477158547753-539'); if (document.body.className.includes('dark-theme')) { iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2101885477158547753&theme=dark" }

jev-skill-suggestion removes that listing and puts a small decision model in front of every prompt instead. The model picks at most one skill, or none, and only that skill’s SKILL.md reaches the context window.

What Jev is

Jev is TypeSafe’s System One decision model. It does not write text. You give it a state and a set of typed questions, and it answers with calibrated probabilities in a few hundred milliseconds. Three primitives cover everything it does:

  • noul: a yes or no question, answered with the probability that the answer is yes
  • choice: pick one option out of a labeled set, answered with a distribution over all of them
  • score: a graded value on a scale you define

The Mod uses choice to rank your skills and noul to ask whether the prompt needs a skill at all. You do not write any of that yourself. It ships configured.

// Detect dark theme var iframe = document.getElementById('tweet-2101485309095514241-450'); if (document.body.className.includes('dark-theme')) { iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2101485309095514241&theme=dark" }

Install it

You need Claude Code 2.1.278 or newer and the function hooks flag.

export CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1
npx claude-code-templates@latest --mod productivity/jev-skill-suggestion

That is the whole install. With no API key it falls back to the engine’s own classifier, so it does something useful before you sign up for anything. Add a TypeSafe key or a Vercel AI Gateway key later, when you want the better decision.

Then run /jev-skill-suggestion:setup. It reads your installed skills, shows you the plan, asks before it touches anything, and writes a backup you can restore. After that, /context should count your skills at 0.

The settings worth touching

Everything has a working default. These are the ones you might actually change:

  • alwaysListed: comma separated skills that stay visible to the model whatever the prompt. Use it for the one or two you always want in reach.
  • neverSuggested: skills Jev is never offered. They still run when you type /name.
  • gateThreshold: how sure the model has to be that the prompt needs a skill at all. Default 0.3. Raise it if you get suggestions on questions that were only questions.
  • timeoutMs: the latency budget, default 800. Past it, the prompt goes through with no suggestion.

They go in pluginConfigs in your settings, or through /config:

{
  "pluginConfigs": {
    "jev-skill-suggestion": {
      "alwaysListed": "commit-helper",
      "gateThreshold": 0.45,
      "timeoutMs": 800
    }
  }
}

Every failure path is fail-open. If a request errors or runs past the budget, the prompt goes through untouched. It never blocks you.

Is the trade worth it

I benchmarked it against Claude models doing the picking instead: 26 skills, 10 prompts, three runs each.

  • Jev: 90% accuracy, 330ms median
  • Haiku: 80%, 1.24s
  • Sonnet, Opus and Fable: 100%, between 2.2s and 4.5s

// Detect dark theme var iframe = document.getElementById('tweet-2102167460136796594-439'); if (document.body.className.includes('dark-theme')) { iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2102167460136796594&theme=dark" }

So it is not the most accurate option. It is 4 to 14 times faster than asking a model, and it is the only one of the five that also keeps the skill list out of the context window, which was the point. With a handful of skills none of this matters. Past twenty, the listing is real money on every session.

Where the Mods live

Both Jev Mods are on aitmpl.com, with the install command, the full options table and the source for each one. This article covers jev-skill-suggestion. Its companion, jev-model-router, puts the same decision model in front of the model and effort choice on every turn, and the two run together.

The rest of the catalog, Mods, Agents, Skills, Commands, Hooks, MCPs and Settings, installs the same way from aitmpl.com.

You can follow me on my social media:

Share