catchup

name: catchup description: Generate a research onboarding document for the current Open-FARS project. Explains the field, survey findings, our innovation, hypotheses, experiment design, and current results in accessible language. Persisted to project catchup/ directory. user_invocable: true

Open-FARS Research Catch-Up

Overview

Generate an accessible, structured research onboarding document for the current Open-FARS project. Designed for someone who needs to quickly understand what this research is about, why it matters, what we're doing, and where we are — without prior domain knowledge.

Each run produces a self-contained document that reflects the latest project state: new experiment results, updated hypotheses, revised plans. Prior catch-up documents are preserved for history.

Usage

/catchup                              # Full catch-up, auto-detect project
/catchup 重点讲 SAE 和电路              # Full catch-up with emphasis area
/catchup email                        # Full catch-up + send via email-notify
/catchup email 给新加入的同学看          # Email catch-up with context

Parsing rules

1. If arguments contain email → generate document AND send via email-notify
2. Remaining text after email keyword → user focus / emphasis area
3. If no email keyword, all text is user focus

Data Collection Steps

Collect ALL of the following before generating the document. Use parallel tool calls where possible.

Step 1: Project Metadata

Read .open-fars/meta/registry.yaml:

• Direction slug, project slug, paper title, target venue
• Current pipeline stage statuses
• Known degradations

Step 2: Survey Foundation

Read from .open-fars/survey/{direction}/:

• gaps.md — research gaps (the "why" of this project)
• INDEX.md — paper count, categories
• literature-network.md — key citation chains (if exists)
• Scan papers/ directory for count and topic coverage

Extract:

• Total papers surveyed
• Top 3-5 gaps that motivated this project
• Key prior work chain (the 5-8 papers that form the intellectual lineage)

Step 3: Ideation Context

Read from .open-fars/ideation/{direction}/:

• INDEX.md — how many ideas were generated, which was selected
• The selected idea file (e.g., ideas/idea-01-*.md) — core proposal, novelty check

Extract:

• Why this idea scored highest
• What makes it novel (the novelty check results)
• The core insight in one sentence

Step 4: Research Plan

Read from .open-fars/plan/{direction}/{project}/LATEST.md:

• Research questions (RQ1-RQ4)
• Hypotheses (H1-H4) with thresholds
• Model and task selection rationale
• Experiment design (E1-E6) overview
• Ablation design (A1-A5) overview

Step 5: Current Results

Read latest experiment result JSON files from .open-fars/projects/{project}/experiments/results/:

• For each completed/in-progress experiment, extract headline numbers
• Compare results against hypothesis thresholds from the plan
• Note any surprises or failures

Step 6: Prior Catch-Up Documents

ls .open-fars/projects/{project}/catchup/ 2>/dev/null

If prior documents exist, read the most recent one to understand what has changed since last time. The new document should note "what's new since last catch-up" if a prior exists.

Document Structure

Generate the document in the following structure. Use Chinese for section headers and narrative (user's primary language for communication), keep technical terms in English.

# {project_title} — 研究入门与进展串讲

**项目**: {paper_title}
**目标会议**: {target_venue}
**日期**: {today}
**方向**: {direction description}

---

## 一、一句话概括

{用一句通俗的话说清楚这个研究在干什么，不用任何术语}

## 二、从日常经验出发

{用 2-3 个具体例子，让没有 ML 背景的人也能理解这个问题的直觉}

## 三、研究领域概览

{画出领域的两个(或多个)社区，说明我们在它们之间的位置}
{说明为什么这个交叉点是空白的}

### 文献调研概况

- 调研论文数: {N}
- 核心论文链: {5-8 篇关键论文，按逻辑链排列，每篇一句话说明它贡献了什么}
- 研究空白: {top 3 gaps，每个用 2-3 句话解释}

## 四、关键术语表

{按逻辑顺序（不是字母顺序）列出所有关键术语}
{每个术语: 一句定义 + 一个类比或例子}

| 术语 | 解释 | 类比/例子 |
|------|------|----------|
| ... | ... | ... |

## 五、我们的创新点

{3 个创新点，每个:}
{1. 一句话概括}
{2. 为什么之前没人做}
{3. 为什么它重要}

## 六、研究设计

### 假设

{每个假设用「通俗版」+ 「技术版」两种方式表述}
{标注当前验证状态: ✅ 通过 / ❌ 失败 / 🔄 待验证}

### 实验一览

| 实验 | 做什么 | 回答哪个问题 | 当前状态 | 关键结果 |
|------|--------|-------------|---------|---------|
| E1 | ... | RQ1 / H1 | ... | ... |
| ... | ... | ... | ... | ... |

### 任务选择

{为什么选这 5 个任务，每个任务一句话说明}

### 模型选择

{为什么选这些模型，各自的角色}

## 七、当前实验结果

{对每个已有结果的实验:}
{- 关键数字}
{- 对照假设阈值的判断}
{- 意味着什么（通俗解释）}

{如果有假设失败的情况，诚实说明并讨论可能原因}

## 八、降级与风险

{当前存在的降级，用通俗语言解释每个降级为什么重要}
{最大的风险是什么}

## 九、数据流与实验依赖图

{用 ASCII 图画出 E1→E2→E3, E1→E4 等依赖关系}
{标注当前进度}

## 十、下一步

{接下来要做什么，优先级排序}

{If prior_catchup_exists: "\n## 附：本次更新内容\n\n{与上次 catch-up 相比有什么新进展}"}

{If user_focus_query: "\n## 重点补充: {user_focus_query}\n\n{针对用户关注领域的深入讨论}"}

Document Persistence

Every catch-up document must be saved to disk:

1. Generate timestamp: TZ=Asia/Shanghai date +"%Y-%m-%d_%H%M"
2. Save to: .open-fars/projects/{project}/catchup/{timestamp}_catchup.md
3. Directory: .open-fars/projects/{project}/catchup/ (create if not exists)
4. Files are append-only — never delete or overwrite prior documents
5. After saving, print the file path to the user

Email Mode

If email is in the arguments, after saving the document:

1. Send via email-notify:

   python3 .claude/skills/email-notify/send-email.py \
     --subject "[Claude Code] Open-FARS 研究串讲: {paper_title_short}" \
     --body-file .open-fars/projects/{project}/catchup/{timestamp}_catchup.md
   

Important

• Accessible first: the primary audience is someone who does NOT know this field. Jargon must be explained on first use.
• Grounded in data: every claim must reference actual files (survey papers, plan, results). Do not fabricate or extrapolate.
• Honest about failures: if a hypothesis failed, say so. If data is insufficient, say so.
• Evolving document: each run captures the project state at that moment. Compare against prior catch-ups when available.
• Terminology consistency: use the same English terms as the paper and plan. Do not translate technical terms into Chinese — keep them in English with Chinese explanation.
• Prior catch-up diff: if a prior catch-up exists, include a "what's new" section highlighting changes in results, status, or understanding.