sawz/Agent-Reach
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Initial: forked from runesleo/x-reader (MIT License) - thank you @runes_leo!

Browse Source
This commit is contained in:
Panniantong
2026-02-24 03:00:05 +01:00
commit ee2ad83b12
25 changed files with 2512 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
skills/analyzer/skill.md
+131
View File
@@ -0,0 +1,131 @@
# Content Analyzer Skill
> Any content → structured analysis report with actionable insights
## Trigger
When user sends content (URL, text, or transcript) with analysis intent:
- `/analyze [URL]`
- "Analyze this article"
- "What are the key takeaways?"
- Auto-triggered after video/podcast transcription (from video skill)
## Pipeline
### Step 1: Get Content
Choose tool based on input type:
| Input | Tool |
|-------|------|
| Tweet URL | `fetch_tweet` or Jina Reader |
| Web URL | `WebFetch` or Jina Reader |
| Local file | Read file directly |
| Transcript from video skill | Use directly |
### Step 2: Multi-Dimensional Analysis
Scan content across these dimensions. Only output dimensions with actual content — skip empty ones.
```markdown
## 📖 Summary
[1-3 sentence core thesis]
**Source**: [author/publisher] · [date]
**Type**: [tweet/article/video/podcast/report]
---
## 💡 Key Insights
### 🎯 Core Arguments
- **Thesis**: [Main argument or finding]
- **Evidence**: [Supporting data or reasoning]
- **Strength**: [How convincing? What's missing?]
### 🤖 Tools & Methods
- **What**: [Tools, frameworks, or techniques mentioned]
- **How**: [How they're used or applied]
- **Relevance**: [Could you use this?]
### ⚙️ Workflow Ideas
- **Optimization**: [Process improvements mentioned]
- **Automation**: [What could be automated]
- **Integration**: [How to fit into existing workflow]
### 📊 Data & Numbers
- **Key metrics**: [Important numbers mentioned]
- **Trends**: [Patterns in the data]
- **Gaps**: [What data is missing]
### ⚠️ Risks & Warnings
- **Pitfalls**: [Explicitly mentioned risks]
- **Blind spots**: [What the author might be missing]
- **Counter-arguments**: [Alternative perspectives]
### 🔗 Resources
- **Tools/APIs**: [Mentioned tools or data sources]
- **People**: [Worth following or referencing]
- **Further reading**: [Related content]
### 🧠 Mental Model Shifts
- **Before**: [Common assumption]
- **After**: [New understanding from this content]
- **Impact**: [How this changes decisions]
---
## ✅ Action Items
### Quick Wins (< 30 min)
- [ ] [Action 1] — Impact: ★★★★ | Effort: Easy
- [ ] [Action 2] — Impact: ★★★ | Effort: Easy
### Deeper Work (1-3 hours)
- [ ] [Action 3] — Impact: ★★★ | Effort: Medium
- [ ] [Action 4] — Impact: ★★ | Effort: Medium
### Exploration (needs validation)
- [ ] [Action 5] — Impact: ★★★ | Effort: Hard | Nature: Exploratory
```
### Step 3: Personalized Relevance (Customizable)
Map insights to YOUR context. Edit the dimensions below to match your own projects, interests, and systems.
```markdown
## 🔄 How This Applies to Me
### My Projects
- **[Project A]**: [How this insight connects]
- **[Project B]**: [What I could apply]
### My Knowledge Base
- **Update**: [Which notes/docs to update]
- **New entry**: [What to add to my knowledge system]
### My Decision Log
- **Changed my mind about**: [what and why]
- **Confirmed my belief that**: [what]
```
> **Customization**: Edit the dimensions in Step 2 and Step 3 to match your own
> domain. A trader might add "Market Impact" and "Risk Assessment". A developer
> might add "Architecture Patterns" and "Tech Debt". Make it yours.
## Output Modes
| Mode | Trigger | Output |
|------|---------|--------|
| **Full** (default) | `/analyze [URL]` | All dimensions |
| **Sparse** | `/analyze [URL] --sparse` | Only hit dimensions, skip empty |
| **Brief** | `/analyze [URL] --brief` | Action items only |
## Best Practices
1. **Scan all dimensions, but don't force-fill** — skip empty dimensions cleanly
2. **Actions must be specific** — not "learn about X" but "read X docs chapter Y"
3. **Distinguish fact from opinion** — mark the author's claims vs verified facts
4. **Source everything** — tag where each insight comes from in the original content
5. **ROI awareness** — not every action is worth doing, assess effort vs impact
skills/video/skill.md
+292
View File
@@ -0,0 +1,292 @@
# Video & Podcast Digest Skill
> Send a video/podcast link → get full transcript + structured summary
## Supported Platforms
| Platform | Type | Subtitles | Whisper Transcription |
|----------|------|-----------|----------------------|
| YouTube | Video | ✅ | ✅ |
| Bilibili | Video | ✅ | ✅ |
| X/Twitter | Video | ❌ | ✅ |
| Xiaoyuzhou (小宇宙) | Podcast | ❌ | ✅ |
| Apple Podcasts | Podcast | ❌ | ✅ |
| Direct links (mp3/mp4/m3u8) | Any | ❌ | ✅ |
## Trigger
Auto-triggered when a media URL is detected:
- YouTube: `youtube.com`, `youtu.be`
- Bilibili: `bilibili.com`, `b23.tv`
- X/Twitter: `x.com`, `twitter.com` (tweets with video)
- Xiaoyuzhou: `xiaoyuzhoufm.com`
- Apple Podcasts: `podcasts.apple.com`
- Direct: `.mp3`, `.mp4`, `.m3u8`, `.m4a`, `.webm`
## Pipeline
### Step 0: Detect Media Type
| URL Pattern | Type | Pipeline |
|-------------|------|----------|
| `xiaoyuzhoufm.com/episode/` | Podcast | → Step 1b (Xiaoyuzhou) |
| `podcasts.apple.com` | Podcast | → Step 1c (Apple) |
| `bilibili.com`, `b23.tv` | Video | → Step 1d (Bilibili API) |
| `.mp3`, `.m4a` direct link | Audio | → Step 2b (direct download) |
| Other | Video | → Step 1a (subtitle extraction) |
### Step 1a: Video — Extract Subtitles
```bash
# Clean up temp files
rm -f /tmp/media_sub*.vtt /tmp/media_audio.mp3 /tmp/media_transcript*.json /tmp/media_segment_*.mp3 2>/dev/null || true
# YouTube (prefer English, fallback Chinese)
yt-dlp --skip-download --write-auto-sub --sub-lang "en,zh-Hans" -o "/tmp/media_sub" "VIDEO_URL"
# Bilibili
yt-dlp --skip-download --write-auto-sub --sub-lang "zh-Hans,zh" -o "/tmp/media_sub" "VIDEO_URL"
```
Check for subtitles:
```bash
ls /tmp/media_sub*.vtt 2>/dev/null
```
- **Has subtitles** → Read VTT content, skip to Step 3
- **No subtitles** → Step 2a (download audio)
### Step 1b: Xiaoyuzhou (小宇宙) — Extract Audio URL
```bash
# Extract CDN direct link from __NEXT_DATA__
# Xiaoyuzhou is a Next.js SPA, but initial HTML contains audio URL in __NEXT_DATA__
AUDIO_URL=$(curl -sL -H "User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36" \
"EPISODE_URL" \
| grep -oE 'https://media\.xyzcdn\.net/[^"]+\.(m4a|mp3)' \
| head -1)
echo "Audio URL: $AUDIO_URL"
# Download audio
curl -L -o /tmp/media_audio.mp3 "$AUDIO_URL"
```
> If curl extraction is empty (rare), fallback: use Puppeteer/browser to get rendered page and extract.
→ Step 2b (check size & transcribe)
### Step 1c: Apple Podcasts — via yt-dlp
```bash
yt-dlp -f "ba[ext=m4a]/ba/b" --extract-audio --audio-format mp3 --audio-quality 5 \
-o "/tmp/media_audio.%(ext)s" "APPLE_PODCAST_URL"
```
→ Step 2b (check size & transcribe)
### Step 1d: Bilibili — API Direct Audio Stream
yt-dlp returns 412 for Bilibili even with cookies. Use Bilibili's API instead:
```bash
# 1. Extract BV number from URL
BV="BV1xxxxx" # Replace with actual BV number
# 2. Get video info (title, duration, CID)
curl -s "https://api.bilibili.com/x/web-interface/view?bvid=$BV" \
-H "User-Agent: Mozilla/5.0" -H "Referer: https://www.bilibili.com/" \
| python3 -c "import json,sys; d=json.load(sys.stdin)['data']; print(f\"Title: {d['title']}\nDuration: {d['duration']}s\nCID: {d['cid']}\")"
# 3. Get audio stream URL
CID=<CID from previous step>
AUDIO_URL=$(curl -s "https://api.bilibili.com/x/player/playurl?bvid=$BV&cid=$CID&fnval=16&qn=64" \
-H "User-Agent: Mozilla/5.0" -H "Referer: https://www.bilibili.com/" \
| python3 -c "import json,sys; print(json.load(sys.stdin)['data']['dash']['audio'][0]['baseUrl'])")
# 4. Download audio (Referer header required, otherwise 403)
curl -L -o /tmp/media_audio.m4s \
-H "User-Agent: Mozilla/5.0" -H "Referer: https://www.bilibili.com/" "$AUDIO_URL"
# 5. Convert to mp3
ffmpeg -y -i /tmp/media_audio.m4s -acodec libmp3lame -q:a 5 /tmp/media_audio.mp3
```
→ Step 2b (check size & transcribe)
### Step 2a: Video — Download Audio (when no subtitles)
```bash
# YouTube may need --cookies-from-browser chrome to bypass bot detection
yt-dlp --cookies-from-browser chrome -f "ba[ext=m4a]/ba/b" --extract-audio --audio-format mp3 --audio-quality 5 \
-o "/tmp/media_audio.%(ext)s" "VIDEO_URL"
```
### Step 2b: Check Audio Size & Segment
```bash
FILE_SIZE=$(stat -f%z /tmp/media_audio.* 2>/dev/null || stat -c%s /tmp/media_audio.* 2>/dev/null)
echo "File size: $FILE_SIZE bytes"
```
- **≤ 25MB (25000000)** → Step 2c (transcribe directly)
- **> 25MB** → Split first, then transcribe each segment
**Splitting large audio (>25MB)**:
```bash
# Get total duration
DURATION=$(ffprobe -v error -show_entries format=duration -of csv=p=0 /tmp/media_audio.* | head -1)
# Split into 10-minute segments (keeps each under 25MB)
SEGMENT_SEC=600
SEGMENTS=$(python3 -c "import math; print(math.ceil(float('$DURATION')/$SEGMENT_SEC))")
# Cut segments
for i in $(seq 0 $((SEGMENTS-1))); do
START=$((i * SEGMENT_SEC))
ffmpeg -y -i /tmp/media_audio.* -ss $START -t $SEGMENT_SEC -acodec libmp3lame -q:a 5 \
"/tmp/media_segment_${i}.mp3" 2>/dev/null
done
```
→ Call Step 2c for each segment **sequentially** (parallel triggers Groq 524 timeout), concatenate results
### Step 2c: Whisper Transcription
**Prerequisite**: `GROQ_API_KEY` environment variable
```bash
# Check API key
if [ -z "$GROQ_API_KEY" ]; then
echo "❌ GROQ_API_KEY not set. Get one at: https://console.groq.com/keys"
exit 1
fi
# Transcribe single file (replace AUDIO_FILE with actual path)
curl -s -X POST "https://api.groq.com/openai/v1/audio/transcriptions" \
-H "Authorization: Bearer $GROQ_API_KEY" \
-H "Content-Type: multipart/form-data" \
-F "file=@AUDIO_FILE" \
-F "model=whisper-large-v3-turbo" \
-F "response_format=verbose_json" \
-F "language=zh" \
> /tmp/media_transcript.json
# Extract plain text
python3 -c "import json; print(json.load(open('/tmp/media_transcript.json'))['text'])"
```
**Whisper model options**:
| Model | Speed | Accuracy | Use Case |
|-------|-------|----------|----------|
| `whisper-large-v3-turbo` | 10x realtime | High | Default choice |
| `whisper-large-v3` | 5x realtime | Highest | Professional/noisy content |
**Language parameter**:
- Chinese: `language=zh`
- English: `language=en`
- Auto-detect: omit language parameter
### Step 3: Structured Summary
Choose output format based on media type:
**Video (≤20 min)**:
1. **Overview** (1-2 sentences)
2. **Key Points** (3-5 bullet points)
3. **Notable Quotes** (if any)
4. **Action Items** (if applicable)
**Podcast (>20 min)**:
1. **Overview** (2-3 sentences: who discussed what)
2. **Chapter Summary** (segmented by topic, 2-3 sentences each)
3. **Key Points** (5-8 bullet points)
4. **Notable Quotes**
5. **Action Items** (if applicable)
## Output Format
### Video
```
## 📺 Video Digest
**Title**: [Video Title]
**Duration**: [x minutes]
**Language**: [Chinese/English]
### Overview
[1-2 sentence summary]
### Key Points
1. [Point 1]
2. [Point 2]
...
### Notable Quotes
> "xxx" — [timestamp]
### Action Items
- [if applicable]
```
### Podcast
```
## 🎙️ Podcast Digest
**Show**: [Podcast Name]
**Episode**: [Episode Title]
**Duration**: [x minutes]
**Guests**: [if any]
### Overview
[2-3 sentences: who discussed what, core conclusions]
### Chapter Summary
#### 1. [Topic] (~xx:xx-xx:xx)
[2-3 sentences of core content]
#### 2. [Topic] (~xx:xx-xx:xx)
[2-3 sentences of core content]
...
### Key Points
1. [Point 1]
2. [Point 2]
...
### Notable Quotes
> "xxx"
### Action Items
- [if applicable]
```
## Error Handling
| Situation | Action |
|-----------|--------|
| No subtitles + no GROQ_API_KEY | Prompt user to set API key |
| No subtitles + has API key | Auto Whisper transcription |
| Xiaoyuzhou curl extraction empty | Use Puppeteer/browser to get rendered HTML |
| Audio >25MB | ffmpeg segment (10min/segment), transcribe sequentially |
| Podcast >2 hours | Warn user about duration, confirm before proceeding |
| Groq 524 timeout | Do NOT parallelize — transcribe sequentially, sleep 5-8s between segments |
| Groq 429 rate limit | 7200s/hour limit, wait for retry-after header, then retry |
| yt-dlp Bilibili 412 | Use Bilibili API instead (Step 1d) |
| yt-dlp YouTube bot detection | Add `--cookies-from-browser chrome` |
| Network timeout | Retry once |
| Spotify links | Inform user: not supported (DRM protected) |
## Groq Whisper Limits
- Max 25MB per request
- Free tier: 7200 seconds of audio/hour (rolling window), ~20 hours/day
- Supported formats: mp3, mp4, mpeg, mpga, m4a, wav, webm
## Dependencies
- `yt-dlp`: video download + subtitle extraction
- `ffmpeg`: audio conversion + segmentation
- `curl`: Xiaoyuzhou audio download, Bilibili API
- `GROQ_API_KEY`: Whisper transcription API (free at https://console.groq.com/keys)