Warning: Unknown data sources detected:
- data/messaging/telegram/ (not in registry)
- data/raw/linear/ (not in registry)

To add support, document the format in source-formats.md and add to sources/_registry.json

Continue processing known sources, but warn about unknown ones.

Execution Steps (FOLLOW EXACTLY)

Step 1: Load State (EXACT)

Read `.claude/state/timeline-state.json`
If file doesn't exist → treat all files as new
Parse JSON → store in memory as `state`

Step 2: Discover Sources (EXACT)

Run: Glob("data/transcripts/hyprnote/*/_meta.json") → hyprnote_sessions
Run: Glob("data/transcripts/notion/*.md") → notion_files (exclude readme.md)
Run: Glob("data/messaging/whatsapp/*/_chat.txt") → whatsapp_chats
Run: Glob("data/messaging/hailer/*.md") → hailer_files (exclude readme.md)

Step 3: Build Unified Source List (EXACT)

For each hyprnote session:
  - Read _meta.json → extract created_at
  - Add to sources: {type: "hyprnote", date: created_at, path: folder_path}

For each notion file:
  - Extract date from filename (YYYY-MM-DDTHH_MM_SSZ.md)
  - Add to sources: {type: "notion", date: extracted_date, path: file_path}

For whatsapp:
  - Add to sources: {type: "whatsapp", date: "continuous", path: folder_path}

For hailer:
  - Add to sources: {type: "hailer", date: "continuous", path: folder_path}

Sort sources by date DESCENDING (newest first)

Step 4: Identify Changes (EXACT)

For each source in sorted list:
  If type == "hyprnote" or type == "notion":
    If path NOT in state.processed_files[type] → mark as NEW
    Else if file mtime > state.processed_files[type][path].processed_at → mark as CHANGED

  If type == "whatsapp":
    Read first 100 chars of _chat.txt
    If chars != state.whatsapp.file_hash → mark as CHANGED (re-process ALL)

  If type == "hailer":
    For each file in hailer_files:
      Read first 100 chars
      If chars != state.hailer.files[path].file_hash → mark as CHANGED (re-process ALL)

Step 5: Process Each Source (PRESCRIPTIVE + INTERPRETIVE)

For Hyprnote (EXACT steps, then INTERPRET):

1. Read _meta.json → extract title, created_at
2. Read _transcript.json → extract transcripts[0].words array
3. Reconstruct text: words.map(w => w.text).join('').trim()
4. Group words by channel field (speaker separation)
5. Check if _memo.md exists → if yes, read it
6. [INTERPRET] Write 2-4 sentence summary using transcript + memo
7. [INTERPRET] Extract decisions, action items, key insights
8. [INTERPRET] Identify participants and roles
9. [INTERPRET] Rate relevance (High/Medium/Low)

For Notion (EXACT steps, then INTERPRET):

1. Read entire .md file
2. [INTERPRET] Identify speakers from text patterns
3. [INTERPRET] Write 2-4 sentence summary
4. [INTERPRET] Extract decisions, action items, key insights
5. [INTERPRET] Identify participants and roles
6. [INTERPRET] Rate relevance

For WhatsApp (EXACT steps, then INTERPRET):

1. Read entire _chat.txt
2. Parse messages: regex [MM/DD/YY, HH:MM:SS] Sender: Message
3. List all attachments in folder
4. For each image attachment → Read image, describe visual content
5. For each document (.pdf, .md, .docx) → Read and summarize
6. [INTERPRET] Group messages into topic clusters (time gaps >24h, topic shifts)
7. [INTERPRET] For each cluster: write summary, list key events
8. [INTERPRET] Rate relevance per cluster

For Hailer (EXACT steps, then INTERPRET):

1. Read ALL .md files in folder
2. Parse message format (Username, timestamp, content)
3. Combine into chronological order
4. [INTERPRET] Group into topic clusters
5. [INTERPRET] Write summary, extract key points
6. [INTERPRET] Identify participants
7. [INTERPRET] Rate relevance

Step 6: Generate Output (EXACT)

1. Sort all processed entries by date ASCENDING (oldest first for output)
2. Group by month
3. Write to analysis/events-chronological.md using output-template.md format
4. Include YAML frontmatter with counts

Step 7: Validate Output (EXACT CHECKLIST)

Before finishing, verify each entry has:
- [ ] Substantive summary (not just "Meeting occurred" or title repeat)
- [ ] At least one of: decision, action item, or key insight
- [ ] Participants listed
- [ ] Relevance rating

For Hyprnote with _memo.md:
- [ ] Memo content incorporated in entry

For WhatsApp:
- [ ] All images have visual descriptions
- [ ] All documents have content summaries
- [ ] Clusters have date ranges

If any check fails → fix before proceeding

Step 8: Update State (EXACT)

Update state.last_updated = current ISO timestamp
For each processed source:
  Add/update entry in state.processed_files with:
    - processed_at: current timestamp
    - deep_processed: true
    - summary: first 100 chars of generated summary
Write state to .claude/state/timeline-state.json

Supporting Files

Processing Rules - Deep understanding guidelines
Source Formats - Parsing reference
Output Template - Entry format specification

update-timeline

Update Timeline Skill

Source Registry

Loading Sources

Unknown Source Detection