Community코딩 & 개발github.com

Yannyehao/codex-skills

최신 zAI 결과물과 실시간 후보 추적 데이터베이스에서 Yehao의 전체 미국 주식 리뷰를 완성합니다. $us-market-review, 美股复盘, 오늘/이번 주 미국 주식, 오늘 선택된 미국 주식 확인 등과 같은 요청에 트리거됩니다.

codex-skills란 무엇인가요?

이 스킬은 사용자 Yehao의 전체 미국 주식 시장 리뷰를 자동화합니다. 최신 zAI 인공지능 산출물과 실시간으로 업데이트되는 후보 주식 데이터베이스를 기반으로 포괄적인 분석을 제공합니다. 주로 Claude Code나 Codex와 같은 AI 에이전트 환경에서 실행되며, 사용자가 특정 명령어(예: "美股复盘", "오늘의 미국 주식")를 입력하면 트리거됩니다. 스킬은 먼저 참조 문서와 설정 파일을 확인하고, 사전 정의된 감사 스크립트를 실행하여 리뷰의 정확성을 보장합니다. 그런 다음 실시간 데이터베이스에서 현재 월의 모든 데이터를 쿼리하여 오늘의 선택, 활성 관심 종목, 상태 변경 사항, 그리고 월간 원장의 네 가지 명확한 집합을 구성합니다. 각 종목에 대해 실제 차트를 확인하고, 포지션, 추세선, 거래량 등 핵심 기술적 지표를 평가합니다. 최종 결과물은 실행 가능한 결론, 각 종목에 대한 간결한 평가, 상태 변경 내역, 완전한 월간 테이블, 그리고 다음 날을 위한 체크리스트를 포함하여 보고서 형식으로 제공됩니다. 이 스킬은 읽기 전용 작업이며, 데이터베이스 업데이트나 주문 실행은 사용자의 별도 명시적 요청이 있을 때만 수행됩니다.

지원 대상~Claude Code✓Codex CLI~Cursor
npx skills add https://github.com/Yannyehao/codex-skills/tree/HEAD/skills/us-market-review

Installed? Explore more 코딩 & 개발 skills: steipete/bluebubbles, steipete/eightctl, steipete/blucli · View all 6 →

즐겨 사용하는 AI에게 물어보기

이 에이전트 스킬이 미리 로드된 새 채팅을 엽니다.

문서

U.S. market review

Treat "美股复盘" as a full review, not a synonym for reading today's shortlist. The deliverable must reconcile the latest machine output with the live human-maintained candidate database and preserve continuity across days.

Before reviewing

Read references/review-contract.md completely. Read config.local.json beside this file for the local zAI project and private database targets. If it is missing or still contains placeholders, state which live source is unavailable; never guess a database or silently substitute an old snapshot.

Run the bundled artifact audit before drawing conclusions:

& <configured-python> -X utf8 scripts\audit_us_review.py

Use the Python interpreter required by the current machine instructions; the placeholder is not a literal executable name. When invoked outside the installed skill folder, pass --config <absolute-config.local.json>. Stop the review if the audit reports a date mismatch, incomplete report, missing image, hash mismatch, missing visual-read evidence, or mismatched image token. Name the exact gap instead of filling it from memory.

Build the review universe

Query the configured live candidate database for the current calendar month with full pagination. Preserve every selection date for symbols selected more than once.

Form four explicit sets:

  1. Today's selections: every symbol in the latest nightly report, including non-B1 candidates without a B1 chart.
  2. Active watches: every database row whose current human status is 重点关注, 继续观察, or unresolved 待复盘.
  3. Status changes: additions, upgrades, downgrades, eliminations, invalidated plans, and re-selections since the previous review.
  4. Monthly ledger: every symbol selected during the month, including 不再观察 rows.

Use the human 当前状态 as authoritative over a stale machine conclusion. Reconcile the sets before delivery. A symbol present in the database's active set but absent from today's carried list is a coverage conflict, not evidence that it stopped being active.

Read the charts

Actually open every current-day chart for today's selections and every active watch. Check position, white/yellow lines, N structure, pullback volume, J/KDJ, prior distribution, and whether a multi-wave move is already too mature. Do not call file existence, JSON text, price change, or an earlier chart "看过图".

If an active watch is absent from today's report, render its latest daily/weekly chart from current local bars with the project's established renderer into temporary output, then open it. Do not modify production reports merely to obtain the missing view. A previously eliminated monthly row does not require a new chart unless it was selected again, its status changed, or the user names it.

For a non-B1 candidate with no standard B1 image, use its available source evidence and say that no B1 chart was produced. Never pretend its chart passed the B1 visual checks.

Deliver one readable report

Lead with the actionable conclusion: execute, keep watching, or no setup worth forcing. Then include:

  • one concise paragraph for every symbol selected today and every active watch: current judgment, the one decisive reason, and one observable waiting condition;
  • an explicit status-change section, including "无变化" when appropriate;
  • a complete monthly table with symbol, all selection dates, selection count, latest close, return since first/latest selection as clearly labelled, and current human status;
  • tomorrow's shortest useful checklist.

Do not invent a "first attention" ranking, equate alphabetical order with priority, or upgrade WAIT without new evidence. A screening return is not trading P&L and does not prove the stock should have been bought. When there is no clean opportunity, say so directly.

Keep backend hashes, raw state-machine fields, and audit detail out of the reader-facing report unless a failure needs diagnosis. The chat answer and any saved review must express the same conclusions; do not create competing versions.

Writes require a separate instruction

Reviewing is read-only. Write back to Notion, change candidate status, or generate a trade ticket only when the user explicitly asks to "写回", "固化", or equivalent. After an authorized write, read the affected records/page back and verify them. Never place orders.

Viewpoint-account candidates

When the configured symbols XDV.TO and NA.TO (the user's BNC) are in scope, keep them in a separate "观点仓候选" section and follow the exact rules in the reference. They do not replace the U.S. review universe, and XDV is never described as cash parking.

관련 스킬