Process Inbox

Process items from an inbox folder into notes, with automatic concept matching and bidirectional linking.

Arguments

$ARGUMENTS - Path and optional flags.

Path: Can be relative or absolute. Resolves from current working directory.

• /vault:process inbox - Process ./inbox/ relative to cwd
• /vault:process ./notes/inbox - Relative path
• /vault:process /Users/me/vault/inbox - Absolute path
• /vault:process - Defaults to ./inbox/

Flags:

• --dry-run - Show what would be created/updated without doing it
• --yes - Execute without confirmation

Examples:

/vault:process                    # Process ./inbox/, ask to confirm
/vault:process inbox --dry-run    # Preview changes only
/vault:process inbox --yes        # Process without confirmation

Two-Pass Architecture

1. First pass: Parse inbox, extract items, create/update notes
2. Second pass: Scan all touched notes for concept matches, add links + backlinks

Process

Step 1: Detect inbox format

Scan the inbox folder and detect what's there:

• Individual .md files → Each file is one item
• Daily log files (e.g., 2025-01-04.md) → Split by ## headings or --- separators
• Single inbox.md → Split by ## headings or --- separators
• Mixed → Handle each appropriately

Skip hidden files and folders (.DS_Store, etc.).

Step 2: Extract and classify items

For each item extracted, classify it:

URL item - Line is primarily a URL:

• Bare URL: https://example.com/article
• Markdown link: [Title](https://example.com/article)
• URL with brief comment: https://... - interesting approach

Text item - Paragraphs of thought/content without a primary URL.

Mixed item - Substantial text with embedded URLs (treat as text, preserve URLs).

Step 3: Process each item

For URL items - Create lightweight reference note:

1. Fetch page metadata (title, description) using WebFetch
2. Create note with extracted metadata:

---
updated: YYYY-MM-DD
tags: [extracted, from, content]
source: url
---

# Page Title

Brief description from meta tags.

[Original](https://example.com/article)

## Related

Filename: kebab-case from title (e.g., how-to-build-voice-apps.md)

For substantial text items (>280 characters) - Create new atomic note:

1. Generate a concise title from the content
2. Infer relevant tags from the content
3. Create note:

---
updated: YYYY-MM-DD
tags: [inferred, topics]
source: inbox
---

# Generated Title

The captured thought/content.

## Related

Filename: kebab-case from generated title.

For short text items (≤280 characters):

1. Search existing notes for concept matches
2. If strong match found → Append to existing note:

## Added YYYY-MM-DD

The short thought that was captured.

And update the note's updated field.

3. If no match → Create new atomic note anyway (don't lose content)

Step 4: Discover and add links (second pass)

For each note that was created or updated:

1. Extract concepts - Key nouns, phrases, topics from the note
2. Search for matches in existing notes:
   • Filename matching (e.g., "voice UI" matches voice-notes-app.md)
   • Tag overlap (shared tags indicate relation)
   • Content scanning (key phrases appear in body)
3. Rank matches - Multiple signals = stronger match
4. Add links - Top ~5 matches, only if confidence is moderate+

Inline links: Where a concept is mentioned naturally in text, add [[note-name]].

Related section: Add discovered connections to ## Related footer:

## Related

- [[discovered-connection-1]]
- [[discovered-connection-2]]

Step 5: Add backlinks

For each link added (A → B), also add the reverse (B → A):

1. Open the target note B
2. Find or create ## Related section
3. Add - [[A]] if not already present
4. Update updated field

Step 6: Archive processed items

Move processed inbox items to _processed/ with date folder:

inbox/
├── _processed/
│   └── 2025-01-04/
│       ├── random-thought.md
│       └── daily-log.md
└── (now empty)

Step 7: Report results

Processed 5 items from inbox:

Created:
  - how-to-build-voice-apps.md (from URL)
  - notification-fatigue.md (new thought)

Appended:
  - voice-notes-app.md (+1 thought)

Links added:
  - 12 new connections across 4 notes
  - 8 backlinks updated

Archived to: inbox/_processed/2025-01-04/

Subagent Strategy

For large inboxes (10+ items), parallelize with subagents:

Task	Model	Why
URL metadata extraction	haiku	Simple web fetches
Content classification	haiku	Straightforward detection
Tag inference	haiku	Basic NLP
Concept extraction	haiku	Keyword extraction
Link discovery/matching	sonnet	Needs nuanced understanding

Launch batch processing as background tasks, then synthesize results.

Error Handling

• If URL fetch fails → Create note with URL only, mark for retry
• If note creation fails → Log error, continue with other items
• If linking pass fails → Notes are still created, just unlinked

Safety

• Never modify notes outside the vault
• Archive originals before deleting from inbox
• Confirm before executing (unless --yes)
• Dry-run available to preview all changes