serenity-research-model

Imported from quantskills/skill-serenity-research-model (SKILL.md). Install upstream with npx skills add quantskills/skill-serenity-research-model. Copyright stays with the author.

{
  "version": 1,
  "task": {
    "placeholder": "请提供 Serenity 公开帖子、导出数据、线程链接或整理笔记，并说明研究目标",
    "required": true
  },
  "fields": [
    {
      "key": "mode",
      "label": "处理阶段",
      "type": "select",
      "default": "full",
      "options": [
        { "value": "full", "label": "完整研究流程" },
        { "value": "extract", "label": "提取信号" },
        { "value": "clean", "label": "清洗与复核" },
        { "value": "evaluate", "label": "评估公开观点" },
        { "value": "report", "label": "生成研究模型" }
      ]
    },
    {
      "key": "focus",
      "label": "研究重点",
      "type": "text",
      "placeholder": "例如：AI、半导体供应链、特定股票或瓶颈主题"
    }
  ],
  "prompt_template": "{{#task}}任务与材料：\n{{task}}\n\n{{/task}}{{#attachments}}用户上传的材料（已放入工作区）：\n{{attachments}}\n\n{{/attachments}}按 {{mode}} 阶段处理 Serenity 的公开材料{{#focus}}，重点研究 {{focus}}{{/focus}}；将帖子拆解为最小可用信号单元，区分事实与推断、主动论点与引用或回顾内容，保留失败、修正和中性样本，在价格数据可用时评估多周期后续表现，并披露资料边界与未经验证的收益主张，输出中文报告。"
}

Serenity Research Model

Use this skill to turn Serenity public materials into a reproducible research model. The goal is not to verify private profits or copy trades; the goal is to reverse-engineer observable reasoning patterns from public posts and test how those public signals behaved afterward.

Core Workflow

1. Collect public materials.

   • Prefer user-provided X exports, saved post URLs, copied threads, archived HTML, CSV, JSON, or manually curated notes.
   • TweetClaw can provide public X/Twitter search, reply, and user-post exports when source URLs, authors, timestamps, and retrieval dates are preserved. Do not import follower lists; public follower or engagement counts are context only.
   • If browsing is needed, gather only public posts and public summaries. Record source URLs and retrieval date.
   • Do not treat screenshots, follower claims, or viral return numbers as verified account performance.

2. Normalize posts into the data contract in references/data_contract.md.

   • Use scripts/serenity_mvp.py extract --posts <file> --out <dir>.
   • Supported input: .csv, .json, .jsonl, .txt, .md.
   • Minimum fields are created_at and text; include url, id, and author when available.
   • For TweetClaw exports, follow references/tweetclaw_import.md and drop credentials, tokens, cookies, request headers, private DMs, drafts, and unpublished account data before running the pipeline.

3. Decompose each post at the smallest useful unit.

   • Ticker or asset mention
   • Theme and subtheme
   • Bottleneck claim
   • Supply-chain role
   • Evidence type
   • Catalyst
   • Time horizon
   • Risk marker
   • Conviction signal
   • Follow-up or revision relationship

4. Clean and review the first-pass extraction.

   • Use scripts/serenity_mvp.py clean --signals <signals.csv> --posts <source_posts.csv> --out <dir>.
   • If only a stats file is available, use --ticker-stats <ticker_stats.txt> as a fallback.
   • The cleaner filters ticker candidates against the source cashtag whitelist, preserves theme-only rows, and creates review files for quotes, disclaimers, sarcasm-like language, and no-ticker thematic posts.
   • After manually reviewing priority rows, use scripts/serenity_mvp.py auto-review --signals <manual_review_queue.csv> --out <dir> to apply learned first-pass semantic labels such as quote-only, retrospective track record, crowdsourced watchlist, broad basket, active thesis, and post-event explainer.

5. Evaluate public calls only when price data is available.

   • Use scripts/serenity_mvp.py evaluate --signals <signals.csv> --prices <price_dir> --out <dir>.
   • Price files must be one CSV per ticker, named <TICKER>.csv, with date and close.
   • Report 1, 5, 20, 60, and 120 trading-day forward returns plus max drawdown where possible.

6. Build the research model report.

   • Use scripts/serenity_mvp.py report --signals <signals.csv> --out <dir>.
   • If evaluation output exists, include it in the final interpretation.
   • Use references/serenity_axes.md when converting observations into a model.

Interpretation Rules

• Separate observable facts from inferred logic.
• Mark any claimed returns as unverified unless they come from audited records, broker statements, competition results, or a clearly inspectable portfolio record.
• Distinguish research posts from price-moving posts. A viral post can become part of the catalyst.
• Track failed, stale, and revised ideas with the same care as successful ones.
• Avoid building a model only from winners. Include neutral mentions and follow-up corrections when data exists.
• State the data boundary: which posts were included, which were missing, and whether replies/highlights/quotes were captured.

Output Contract

For a completed MVP run, produce:

• signals.csv: one row per ticker-level public signal.
• theme_summary.csv: aggregated themes, evidence types, and conviction markers.
• manual_review_queue.csv: rows that need human review for quote context, no-ticker themes, disclaimers, or ambiguous rhetoric.
• quote_relationships.csv: rows where the signal depends on quoted text.
• manual_review_priority_top200_reviewed.csv or another *_reviewed.csv: semantic review labels with decision, weight, and reason.
• correction_log.md: documented cleaning rules and counts.
• serenity_model_report.md: concise report with profile, logic tree, evidence map, risk map, and reusable research checklist.
• Optional signal_evaluation.csv: forward-return evaluation from price data.

When Extending

After the Serenity MVP works, use the same schema to compare another trader. Only create a generalized trader-skill generator after at least two trader-specific runs expose stable common fields and failure modes.