/discover

name: discover
description: "Given a Japanese NLP GitHub repo/model/dataset (URL / owner/repo / tool name) OR a topic, find what's already in awesome-japanese-nlp-resources and discover related resources NOT yet listed (contribution candidates). Mines the bundled dataset, then expands via web research across GitHub and Hugging Face. Use when the user names a SPECIFIC repository, model, or tool and wants alternatives/equivalents, OR wants to discover Japanese NLP resources for a topic that are NOT yet in the list, OR wants to prepare a contribution. Trigger phrases include 'mecabに似たツール', 'fugashiの代替', 'alternatives to fugashi', 'repos like manga-ocr', 'what else is like sudachi', 'リストに無い新しい日本語NLP', 'awesome-japanese-nlpに追加できそうな', '最近公開された日本語NLPツール', 'find unlisted Japanese NLP repos', 'new Japanese models on Hugging Face', 'contribute a new resource'. For a simple lookup of what already exists, use the search skill instead."
argument-hint: [github-url | huggingface-url | owner/repo | tool-name | topic]
allowed-tools: Bash WebSearch WebFetch

Discover Japanese NLP resources related to the user's query — both what's already in awesome-japanese-nlp-resources and what's not yet listed.

Claude Code and Codex

This skill is shared by the Claude Code and Codex versions of the plugin. The steps are the same in both tools; only these details differ:

• **Query** — Claude Code: the arguments of `/awesome-japanese-nlp-resources:discover`, appended at the end of this skill as `ARGUMENTS: …`. Codex: the user's message that invoked `$awesome-japanese-nlp-resources:discover`, minus that `$…` mention. If the skill was picked automatically rather than invoked by name, use the user's request as the query.
• **Plugin root** — Claude Code: `${CLAUDE_PLUGIN_ROOT}`. Codex: the directory two levels above this `SKILL.md` (use its absolute path).
• **Shell** — run the commands below with Claude Code's `Bash` tool or Codex's shell tool. Copy each Python script in full and run it as written, changing only its placeholders (`RESOURCES_PATH`, `SEED`, the keyword list, the candidate list) — don't shorten it, drop passes, or alter its scores and thresholds.
• **Web** — Claude Code: `WebSearch` to search and `WebFetch` to read a page. Codex: the built-in web search tool (search, then open the page). Do not use the `gh` CLI, `curl`, or other network commands from the shell.
• **Commands** — write any command you show the user in the current tool's form: `/awesome-japanese-nlp-resources:<skill>` in Claude Code, `$awesome-japanese-nlp-resources:<skill>` in Codex.

The "Already in the list" results and the "not yet in the list" check both come from the bundled data. If the data file can't be read (for example, shell commands are blocked or fail to start), say so and link https://github.com/taishi-i/awesome-japanese-nlp-resources instead of guessing which resources are listed.

Instructions

Preamble — Establish the current date

Before anything else, run this once and remember the values — every step that mentions a year refers to them:

echo "YEAR_NOW=$(date +%Y)"
echo "YEAR_PREV=$(($(date +%Y) - 1))"

Substitute these wherever this skill writes `${YEAR_NOW}` or `${YEAR_PREV}`. **Do not hardcode years.**

Step 0 — Handle empty input

If the query is empty or blank, treat it as a **topic-mode** request for a general search for the latest Japanese NLP resources. Use these defaults for the rest of the steps:

• **Topic label**: "Latest Japanese NLP Resources" (use "最新の日本語NLPリソース" only when the user's query was written in Japanese)
• **Keywords for Step 3 (topic mode)**: `llm`, `bert`, `embed`, `speech`, `morpholog` — short stems for local substring matching (a phrase like `japanese language processing` almost never occurs verbatim in a description and would silently match nothing; see Step 3's substring-matching note)
• **Web searches for Step 6**: focus on recency — add `${YEAR_PREV} ${YEAR_NOW}` to every query, and include:
• `japanese NLP new library github ${YEAR_NOW}`
• `日本語 NLP 新しい ライブラリ github ${YEAR_NOW}`
• `japanese NLP new model huggingface ${YEAR_NOW}`
• `huggingface japanese nlp ${YEAR_PREV} ${YEAR_NOW} new`
• **Output heading**: "Latest Japanese NLP Resources" instead of `Discover: "<query>"`

Then skip Step 1 (mode is topic mode) and continue from Step 2.

Step 1 — Classify the input: seed mode or topic mode

**Seed mode** — the query names ONE specific existing repository, model, or tool:

(a) **A direct identifier** — a full GitHub URL (`https://github.com/owner/repo`), a full Hugging Face URL (`https://huggingface.co/owner/name` or `https://huggingface.co/datasets/owner/name`), an `owner/repo` pair, or a bare tool/model name (e.g. `mecab`). Pass it straight through as `SEED`.

(b) **A natural-language phrase wrapping an identifier** (e.g. `mecabに似たツール`, `alternatives to fugashi`) — extract the core identifier (the URL, `owner/repo`, or tool/model name) and pass only that as `SEED`.

If (a) or (b) applies, proceed to Step 2 in **seed mode**.

**Topic mode** — the query is a descriptive/topical phrase with **no single specific name** in it (e.g. `日本語の要約データセット`, `japanese sentiment analysis dataset`, `形態素解析`). There is no identifier to extract and no reason to force one — proceed to Step 2 in **topic mode** directly. Do not invent a seed; topic mode already covers this case via keyword search plus web discovery.

Keep the full query for the language-detection rule in Step 9.

Step 2 — Locate the data files

The data ships with the plugin at `data/resources.json` under the plugin root (see "Claude Code and Codex" above). Resolve its absolute path, falling back to a scoped search only if the install is unusual:

PLUGIN_ROOT="${CLAUDE_PLUGIN_ROOT}"  # Codex: replace with the plugin root, two levels above this SKILL.md
RESOURCES_PATH="$PLUGIN_ROOT/data/resources.json"
[ -f "$RESOURCES_PATH" ] || RESOURCES_PATH="$(find "${CODEX_HOME:-$HOME/.codex}/plugins" "$