bestagentkits

Jev Skillful — Development skill for Claude Code

Development community

Per-prompt capability router for coding agents: resolves installed skills, MCP servers, agents and commands against your prompt via TypeSafe Jev, and measures whether the injection actually helps.

How to install Jev Skillful

This entry records only its repository, not the path inside it, so there is no exact command to give. Open bestagentkits/jev-skillful and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Jev Skillful does

Per-prompt capability router for coding agents: resolves installed skills, MCP servers, agents and commands against your prompt via TypeSafe Jev, and measures whether the injection actually helps.

Alternatives in Development

Bypass 403 — Probe a 403/401 endpoint with the most-paid bypass tricks (header injection, path encoding, method swap, WAF f 4.5k ★
Obsidian Ingest — Ingest a source into the vault — the vault rewrites itself around new knowledge 277 ★
Benzi — Benzi is compiler-backed code intelligence (Agent+MCP) 102 ★

README

Skillful

A capability router for coding agents. It watches your prompts, decides whether a skill, MCP server, subagent or slash command you already have installed is relevant, and injects at most one suggestion into the agent's context.

npx @mrgoonie/skillful install

You bring your own `TYPESAFE_API_KEY`. There is no server, no account, and nothing is collected.

The problem

The usual way to make an agent aware of a skill is to list every skill in the context window. This does not scale. [Skill Retrieval Augmentation for Agentic AI](https://arxiv.org/abs/2604.24594) found that as the skill corpus grows, context budget is consumed quickly and **the agent becomes less accurate at picking the right skill**. The same work also found that current agents tend to load skills at a similar rate regardless of whether the task actually needs an external capability — so the bottleneck is both *which* capability and *whether* to load one at all.

That is the situation this project targets. A development machine here has around 230 skills across two agent runtimes, plus MCP servers, subagents and commands.

How it works

Your prompt
  │
  ├─ Hook (Claude Code / Codex / Pi / OMP)
  │    cache lookup by prompt hash + catalog fingerprint
  │    hit  → inject (<250ms)
  │    miss → route within a 2000ms budget
  │    error or over budget → inject one reminder line, never fail the prompt
  │
  ├─ Catalog scan (local, no tokens)
  │    every capability on the machine, normalised
  │    BM25 + per-kind quota → a shortlist of about 15
  │
  ├─ One TypeSafe request
  │    a `choice` question over the shortlist plus `none`
  │    per-candidate `noul` questions to rank the runners-up
  │
  └─ Inject at most 1 primary + 2 runner-ups, or nothing

Two design choices are worth stating because they are deliberate:

**The shortlist is what the model sees, not what gets injected.** Injecting all fifteen candidates on every prompt would reproduce the exact "enume