JEV 100 Skill

This folder is a standalone Codex skill. Upload the entire folder to GitHub; SKILL.md is its entry point. It turns a search goal into a source-linked shortlist: choose a source, collect real candidates, ask Jev the same evidence-based questions for each candidate, and return up to five results with verification steps.

What works in this package

• A 104-entry source registry: 100 candidate free-access sources, three reference-only sources, and one optional paid aggregator. This is a routing catalog, not 100 implemented or tested scrapers.
• A working GitHub Search API adapter that collects up to 100 public repositories and README excerpts, then scores each repository with a separate Jev request.
• A separate six-dimension similarity ranker for a supplied, sourced company JSON pool (1–100 companies).
• An optional IT桔子 MCP demonstration that searches a small current company set and ranks the product dimension only when that is the only evidenced dimension. It does not export a 100-company database.

The local JEV 100 webpage and earlier benchmark files are in the separate StartupMirror application and are not needed to use this skill.

Install

Requires Node.js 20 or newer. From this folder:

npm ci
node scripts/doctor.mjs

Set AI_GATEWAY_API_KEY in the process environment before a Jev run. GITHUB_TOKEN is optional for higher GitHub API limits. The IT桔子 demonstration also needs ITJUZI_MCP_API_KEY. Keep keys outside this folder and Git. See setup.

To install as a local Codex skill, clone this repository into your Codex skills directory as jev100-skill (git clone https://github.com/Chloe-JY/jev100 jev100-skill). SKILL.md sits at the repository root, so the clone is the skill folder.

Run a GitHub shortlist

Plan search terms for the user's specific goal, then run:

node scripts/route.mjs "Find an open-source meeting assistant that works offline"
node scripts/github-fetch.mjs --query="offline meeting assistant" --query="local transcription action items" --limit=10 --out=candidates.json
node scripts/rank-jev.mjs --input=candidates.json --out=ranking.json --goal="Find a reusable offline meeting assistant" --must="Can run without cloud inference"

Increase --limit to 100 after checking search relevance, API availability and expected Jev cost. The first command routes source choice; github-fetch.mjs retrieves the actual records. A hard requirement with missing evidence is placed in a verification queue, rather than silently treated as met. Output JSON retains per-candidate scores and status.

Run company similarity

Prepare a UTF-8 JSON file with { "idea": {...}, "companies": [...] } using the schema in company matching. Every company needs a snapshot date, six explicitly present fact fields (unknown = null) and at least one dated source. Then run:

node scripts/company-match.mjs --input=companies.json --out=matches.json

This scores only dimensions with facts on both sides and reports coverage. The corpus is supplied by the user or a separately authorized retrieval step; the package does not bundle 100 verified companies.

For a small live IT桔子 product-only demonstration, after setting both keys:

node scripts/itjuzi-demo.mjs

The script prints a compact ranking and does not save MCP profiles to disk. See company matching for its evidence limits and provider access terms.

Files

• SKILL.md — agent instructions and decision boundaries.
• references/ — source registry, setup, pipeline, company schema and evaluation guidance.
• scripts/ — source routing, checks, GitHub retrieval, Jev ranking and company matching.
• package-lock.json — reproducible MCP client installation.

The .gitignore protects local keys, generated candidate snapshots, rankings and dependencies from an ordinary git add.