summarize

/asciinema-tools:summarize

面向大型 .txt 录制文件的 AI 驱动迭代式深入分析。使用引导式抽样和 AskUserQuestion，逐步探索内容。

自我进化技能：此技能会在使用中不断改进。如果指令有误、参数发生偏移，或需要采用变通方案，请立即修复此文件，不要拖延。仅针对真实且可复现的问题进行更新。

理念

无法完整读取大型录制文件（1GB 以上）。此命令采用以下方式：

1. 初始引导 — 你想查找什么？
2. 策略性抽样 — 开头、中间、结尾，以及针对关键词的片段
3. 迭代式细化 — 使用 AskUserQuestion 深入探查发现
4. 渐进式综合 — 通过多轮分析逐步建立理解

参数

参数	描述
file	.txt 文件路径（由 .cast 转换而来）
--topic	初始关注领域（例如，“ML 训练”、“错误”）
--depth	分析深度：quick、medium、deep
--output	将发现保存到 Markdown 文件

工作流

阶段 1：初始引导

AskUserQuestion:
  question: "What are you trying to understand from this recording?"
  header: "Focus"
  options:
    - label: "General overview"
      description: "What happened in this session? Key activities and outcomes"
    - label: "Key findings/decisions"
      description: "Important discoveries, conclusions, or decisions made"
    - label: "Errors and debugging"
      description: "What went wrong? How was it resolved?"
    - label: "Specific topic"
      description: "I'll specify what I'm looking for"

阶段 2：文件统计信息

/usr/bin/env bash << 'STATS_EOF'
FILE="$1"

echo "=== File Statistics ==="
SIZE=$(ls -lh "$FILE" | awk '{print $5}')
LINES=$(wc -l < "$FILE")
echo "Size: $SIZE"
echo "Lines: $LINES"

echo ""
echo "=== Content Sampling ==="
echo "First 20 lines:"
head -20 "$FILE"

echo ""
echo "Last 20 lines:"
tail -20 "$FILE"

echo ""
echo "=== Keyword Density ==="
echo "Errors/failures:"
grep -c -i "error\|fail\|exception" "$FILE" || echo "0"
echo "Success indicators:"
grep -c -i "success\|complete\|done\|pass" "$FILE" || echo "0"
echo "Key decisions:"
grep -c -i "decision\|chose\|selected\|using" "$FILE" || echo "0"
STATS_EOF

阶段 3：策略性抽样

根据文件大小进行策略性抽样：

对于小于 100MB 的文件：

# Sample head, middle, tail (1000 lines each)
head -1000 "$FILE" > /tmp/sample_head.txt
tail -1000 "$FILE" > /tmp/sample_tail.txt
TOTAL=$(wc -l < "$FILE")
MIDDLE=$((TOTAL / 2))
sed -n "${MIDDLE},$((MIDDLE + 1000))p" "$FILE" > /tmp/sample_middle.txt

对于大于 100MB 的文件：

# Keyword-targeted sampling
grep -B5 -A20 -i "$TOPIC_KEYWORDS" "$FILE" | head -5000 > /tmp/sample_targeted.txt

阶段 4：初步分析

阅读样本并提供初步发现。然后询问：

AskUserQuestion:
  question: "Based on initial analysis, what would you like to explore deeper?"
  header: "Drill down"
  multiSelect: true
  options:
    - label: "Specific timeframe"
      description: "Jump to a particular section (e.g., 'around line 50000')"
    - label: "Follow keyword trail"
      description: "Search for specific patterns and expand context"
    - label: "Error investigation"
      description: "Deep dive into errors and their resolution"
    - label: "Success moments"
      description: "What worked? What were the wins?"
    - label: "Generate summary"
      description: "Synthesize findings into a report"

阶段 5：迭代式深入分析

针对每个选定的关注领域：

1. 提取相关片段，使用带上下文的 grep
2. 阅读并分析提取的内容
3. 向用户报告发现
4. 通过 AskUserQuestion 询问下一步操作

AskUserQuestion:
  question: "Found {N} relevant sections. What next?"
  header: "Continue"
  options:
    - label: "Show me the most significant"
      description: "Display top 3 most relevant excerpts"
    - label: "Search for related patterns"
      description: "Expand search to related keywords"
    - label: "Move on"
      description: "I have enough on this topic"

阶段 6：综合

AskUserQuestion:
  question: "Ready to generate summary. What format?"
  header: "Output"
  options:
    - label: "Concise bullet points"
      description: "Key findings in 10-15 bullets"
    - label: "Detailed markdown report"
      description: "Full report with sections and evidence"
    - label: "Executive summary"
      description: "1-paragraph high-level summary"
    - label: "Save to file"
      description: "Write findings to markdown file"

关键词库

交易／ML 领域

sharpe|drawdown|backtest|overfitting|regime|validation
model|training|loss|epoch|gradient|convergence
feature|indicator|signal|position|portfolio

开发领域

error|exception|fail|bug|fix|debug
commit|push|merge|branch|deploy
test|assert|verify|validate|check

Claude Code 领域

tool|bash|read|write|edit|grep
task|agent|subagent|spawn
permission|approve|reject|block

使用示例

# Interactive exploration
/asciinema-tools:summarize session.txt

# Focused on ML findings
/asciinema-tools:summarize session.txt --topic "ML training results"

# Quick overview
/asciinema-tools:summarize session.txt --depth quick

# Full analysis with report
/asciinema-tools:summarize session.txt --depth deep --output findings.md

输出示例

# Session Summary: research-repo-research_20251226

## Overview

- **Duration**: 4 days (Dec 26-30, 2025)
- **Size**: 12GB recording → 3.2GB text
- **Primary Focus**: ML robustness research

## Key Findings

### 1. Training-Evaluation Mismatch (CRITICAL)

- MSE loss optimizes magnitude, but Sharpe evaluates direction
- Result: 80% Sharpe collapse from 2024 to 2025

### 2. Fishr λ=0.1 Solution (BREAKTHROUGH)

- Gradient variance penalty solves V-REx binary threshold
- Feb'24 Sharpe: -6.14 → +6.14

### 3. Model Rankings

| Model  | Window | Sharpe |
| ------ | ------ | ------ |
| TFT    | 15mo   | 1.02   |
| BiLSTM | 12mo   | 0.50   |

## Evidence Locations

- Line 15234: "Fishr λ=0.1 SOLVES the V-REx binary threshold problem"
- Line 48102: Phase 4 results summary table

## Next Steps Identified

1. TFT 15mo + Fishr training
2. PSR/PBO statistical validation
3. Agent research synthesis

故障排除

问题	原因	解决方案
文件过大	录制文件超出内存限制	使用 --depth quick 仅进行抽样
未找到关键词	领域不符或内容稀疏	尝试更换 --topic 关注领域
抽样超时	文件非常大	延长终端超时时间或使用 grep
grep 上下文错误	缺少 GNU grep	brew install grep（BSD grep 有限制）
未保存输出文件	权限被拒绝	检查 --output 的写入权限

相关命令

• /asciinema-tools:convert - 先将 .cast 转换为 .txt
• /asciinema-tools:analyze - 基于关键词的分析（更快，深度较浅）
• /asciinema-tools:finalize - 处理孤立的录制文件

执行后复盘

此技能完成后，在结束任务前进行复盘：

0. 确认文件位置。 — 编辑前找到此 SKILL.md 的规范路径。
1. 哪里失败了？ — 修复导致问题的指令。
2. 哪些地方比预期更有效？ — 将其提升为推荐实践。
3. 哪些内容发生了偏移？ — 修复与实际情况不再相符的脚本、引用或依赖项。
4. 记录下来。 — 添加进化日志条目，注明触发原因、修复内容和证据。

不要拖延。下一次调用会继承你留下的一切。