line-dev

AI agent skills for LINE platform development. Build bots, login flows, LIFF apps, MINI Apps, notification messages, and LINE stickers/emoji/themes with up-to-date API references.

Contents

• Install as Claude Code Plugin
• Install as Skills
• Available Skills
• Testing

---

Claude Code Plugin

Install all skills as a Claude Code plugin:

/plugin marketplace add abgne/line-dev
/plugin install line-dev@abgne-line-dev

Or test locally without installing:

claude --plugin-dir /path/to/line-dev

After install, reload with /reload-plugins.

---

Install as Skills

All Skills

npx skills add abgne/line-dev

Specific Skills

npx skills add abgne/line-dev@messaging-api
npx skills add abgne/line-dev@line-login
npx skills add abgne/line-dev@line-liff
npx skills add abgne/line-dev@line-mini-app
npx skills add abgne/line-dev@line-notification-message
npx skills add abgne/line-dev@line-creators-market

---

Available Skills

Skill	Description
messaging-api	Webhook, push/reply/multicast, Flex Message, Rich Menu, narrowcast, audience, insights, coupons, channel tokens
line-login	OAuth 2.0, PKCE, ID Token JWT verification, token management, bot linking, login button design
line-liff	LIFF SDK, liff.init, sendMessages, Share Target Picker, QR scan, permanent links, pluggable SDK
line-mini-app	Service Messages, Common Profile Quick Fill, In-App Purchase, Console setup, submission review
line-notification-message	Phone-number-based PNP push (LON), SHA256 hashing, template/flexible types, delivery webhook, consent flow, SMS auth
line-creators-market	Sticker creation (7 types), emoji, themes, technical specs, review guidelines, revenue model, AI usage declaration, LINE Sticker Maker, market strategy

Each skill includes reference files covering API specs, expert guidance, and region-specific details for Japan, Thailand, and Taiwan.

Skills load progressively — metadata is always in context, SKILL.md body loads on trigger, reference files load as needed.

---

Testing Trigger Accuracy

Each skill has an assessment set (should-trigger + should-not-trigger queries in 4 languages: en, ja, zh-TW, th) to measure description quality.

Setup

cd scripts
python3 -m venv .venv
source .venv/bin/activate
pip install claude-agent-sdk

Run

End-to-End (actual Claude Code invocation — requires claude CLI):

# Test one skill
./test_skill_e2e.sh messaging-api --verbose

# Multiple runs for stability
./test_skill_e2e.sh messaging-api --runs 3 --verbose

Simulated (Agent SDK assessment — faster, lower cost):

# Test one skill
./test_skill.sh messaging-api --max-iterations 1 --verbose

# Test all skills
./test_all.sh --verbose

# Auto-optimize description (iterate up to 3 times)
./test_skill.sh line-login --max-iterations 3 --verbose --output results.json

Current Scores

End-to-End (actual claude -p invocation)

Runs each query through Claude Code and checks whether the Skill tool was actually called.

Skill	Accuracy	Queries	Date
line-notification-message	100%	73/73	2026-03-23
line-liff	96%	69/72	2026-03-20
line-login	95%	62/65	2026-03-20
line-mini-app	94%	166/176	2026-03-23
messaging-api	93%	65/70	2026-03-20
line-creators-market	88%	60/68	2026-03-20

Simulated (Agent SDK assessment)

Uses Claude Agent SDK to simulate the triggering decision.

Skill	Accuracy	Queries	Date
line-notification-message	100%	73/73	2026-03-20
line-liff	92%	66/72	2026-03-20
line-login	92%	60/65	2026-03-20
line-mini-app	99%	174/176	2026-03-20
messaging-api	100%	70/70	2026-03-20
line-creators-market	100%	68/68	2026-03-20