¿Qué hace sevensignal?
Fan-built agent skills for BTS fans. English-first, with Korean summaries.
팬이 만든 BTS 팬용 에이전트 스킬 모음입니다. 영어가 기본이며 한국어 요약을 함께 제공합니다.
Status: v0.1.0 is released. Two skills: bts-ticket-safety (complete and tested) and bts-official-news (new, unreleased). Community validation is next.
What this is
A collection of open-source skills that an AI agent can load to help BTS fans with practical, safety-critical tasks. The first skill focuses on the highest-risk moment in fandom life: buying concert tickets without getting scammed.
The name: "seven" is a nod to the seven members. "Signal" is what the skills do - separate trusted signals (official announcements, official sellers) from noise (scams, rumors, resellers). The codename deliberately avoids "BTS" and "ARMY" to keep trademark risk low.
Design principles
- Official sources only. Every trust decision anchors on a deterministic allowlist of official domains, not on vibes or search rankings.
- Deterministic over clever. Domain matching is exact string logic that anyone can audit. No model judgment calls on what "looks official".
- Advise, never transact. These skills help a fan decide. They never buy tickets, hold credentials, or move money.
- Freshness is a feature. Tour dates, onsale times, and prices go stale. Claims that expire are marked for re-verification, with the official place to check.
- Refusal is a valid answer. Out-of-scope requests get a clear refusal plus a safe redirect, not a half-answer.
Repository layout
sevensignal/
LICENSE MIT
README.md this file
DISCLAIMER.md unofficial-project and no-affiliation notice
SECURITY.md how to report security and trust problems
CONTRIBUTING.md content and evidence rules for contributors
CODE_OF_CONDUCT.md contributor covenant
CHANGELOG.md release history
MAINTAINERS.md the weekly maintenance loop and unmaintained policy
.github/ issue forms, pull request template, CI workflows
tools/ executable checks (matcher, validators, eval runner)
skills/
bts-ticket-safety/
SKILL.md the skill definition (English + full Korean guidance)
data/
official-domains.yaml deterministic allowlist of official domains
references/
scam-patterns.md catalog of known ticket-scam patterns
official-channels.md where official announcements actually appear
evals/
README.md eval schema and how to run them
golden.yaml legitimate scenarios -> expected trust verdicts
adversarial.yaml scam scenarios -> expected scam verdicts
refusal.yaml out-of-scope asks -> expected refusals
freshness.yaml claims that expire -> re-verification rules
bts-official-news/
SKILL.md official tour/news tracking, public pages only
evals/ golden, adversarial, refusal, freshness suites
Verify the trust data locally
The domain-matching rules are executable, and the trust data carries its own checks:
python3 tools/test_domain_matcher.py # unit tests (no dependencies)
python3 tools/validate_data.py # schema + 30-day staleness check
python3 tools/eval_runner.py # deterministic eval assertions
python3 tools/check_links.py # live source-link check
See tools/README.md for details. CI runs the first three on every change;
a weekly workflow re-checks data freshness and source links and opens one
stale-data issue when anything needs re-verification.
한국어 요약: 위 명령으로 도메인 일치 규칙과 신뢰 데이터를 직접 검증할 수 있습니다. 데이터가 30일 이상 오래되면 자동으로 경고가 열립니다.
Roadmap (not built yet)
- Fan community safety (Discord/X impersonation detection) - deliberately deferred: impersonation judgment carries real defamation risk and needs more design before it ships as a deterministic skill.
- v0.2.0 after community validation: soft-launch feedback triaged, field
reports turned into fixtures, all
last_verifieddates re-checked.
License
MIT. Home: https://github.com/helenanova/sevensignal