Save

Content Save v2

Routes content to source-specific extraction, then applies common metadata/save pipeline. Uses an isolated .context/<id>/ directory per invocation for file-based handoff between phases.

Step 1: Initialize Context Directory

SAVE_CTX=".context/$(uuidgen | cut -c1-8)"
mkdir -p "$SAVE_CTX"

All intermediate files are written to $SAVE_CTX/. This ensures parallel save invocations never collide.

Step 2: Get Input

Input is provided in $ARGUMENTS.

Parse flags

• --orig flag: If $ARGUMENTS starts with --orig, set keep_original_lang = true and strip --orig (plus any leading whitespace that follows) from $ARGUMENTS.
• Default: keep_original_lang = false

Check input

• If empty (after stripping): Ask the user with AskUserQuestion: "What kind of note should I write from this conversation? (e.g., guide, summary, reference)". Use their answer as user_requirements and skip to Step 5 Conversation path.

Step 3: Detect Input Type

Check in this order:

1. YouTube URL - Contains youtube.com/watch?v=, youtu.be/, youtube.com/shorts/, youtube.com/live/, or m.youtube.com/watch?v=
2. X/Twitter URL - Contains x.com/ or twitter.com/
3. Web URL - Starts with http:// or https:// (not YouTube, not X)
4. Document File - File path ending in .pdf, .docx, .pptx, or .xlsx
5. Conversation Intent - Contains "this conversation", "conversation", or similar phrases. Set user_requirements = $ARGUMENTS and route to Conversation.
6. Raw Text - Everything else

Step 4: Check Duplicate (URL only)

Skip for Document File, Raw Text, and Conversation.

grep -rl "source:.*<DOMAIN_AND_PATH>" --include="*.md" .

If match found:

1. Read matched file's frontmatter for title and path
2. Notify user: "Already saved: [path] ([title])"
3. AskUserQuestion: options ["Overwrite", "Skip"]
   • Overwrite: rm <path>, continue
   • Skip: Abort

Step 5: Phase 1 -- Extract (source-specific)

YouTube

1. Run extraction:

Bash: .claude/skills/save/scripts/extract-defuddle.sh "<URL>"

2. Validate (Change #2): Check exit code AND $SAVE_CTX/extracted.txt non-empty. If failed, inform user and abort.

X Post

Run extraction:

Bash: .claude/skills/save/scripts/extract-defuddle.sh "<URL>"

Validate (Change #2): Check exit code AND $SAVE_CTX/extracted.txt non-empty. If failed, inform user and abort.

Web Page

Run extraction:

Bash: .claude/skills/save/scripts/extract-web.sh "<URL>"

Validate: Check exit code AND $SAVE_CTX/extracted.txt non-empty. If failed, inform user and abort.

Document File

Run extraction:

Bash: .claude/skills/save/scripts/extract-doc.sh "<FILE_PATH>"

Validate: Check exit code AND $SAVE_CTX/extracted.txt non-empty. If failed, inform user and abort.

Raw Text

Write input directly:

1. Write $ARGUMENTS content to $SAVE_CTX/extracted.txt using Write tool
2. Write { "type": "text" } to $SAVE_CTX/meta.json using Write tool

Conversation

Follow .claude/skills/save/conversation.md guide directly (not in subagent):

1. Read conversation.md
2. Execute its workflow with user_requirements
3. Result: $SAVE_CTX/extracted.txt + $SAVE_CTX/meta.json

Step 6: Restructure

Skip for: Raw Text (respect user formatting), Conversation (already markdown), Web Page (already returns structured markdown), Document File (markitdown already returns markdown).

YouTube

1. Read $SAVE_CTX/extracted.txt (the raw transcript).
2. Read .claude/skills/restructure-transcript/SKILL.md and follow its workflow to restructure the transcript into clean prose.
3. Overwrite $SAVE_CTX/extracted.txt with the restructured result using the Write tool.

X Post

1. Read $SAVE_CTX/extracted.txt (raw X post/article text with formatting stripped).
2. Restore markdown formatting:
   • Add heading levels (##, ###) where appropriate
   • Add paragraph breaks between logical sections
   • Format lists, quotes, and emphasis
   • Image/figure references -> italicized captions
   • Do NOT rephrase, summarize, or add content -- structure only
3. Overwrite $SAVE_CTX/extracted.txt with the result using the Write tool.

Step 7: Detect Language + Translate

Skip translation entirely if: keep_original_lang = true OR content is already Korean.

Detect language per type:

• YouTube: Read $SAVE_CTX/meta.json language field.
• X Post: Inspect $SAVE_CTX/extracted.txt first few lines.
• Web Page: Read $SAVE_CTX/meta.json language field. If empty, inspect $SAVE_CTX/extracted.txt first few lines.
• Document File: Inspect $SAVE_CTX/extracted.txt first few lines.
• Raw Text: Inspect content.
• Conversation: Always Korean.

If Korean -> skip. If keep_original_lang = true -> skip. Otherwise -> translate automatically (no user prompt):

YouTube

1. Read $SAVE_CTX/extracted.txt.
2. Read .claude/skills/translate-content/SKILL.md and follow its workflow to translate to Korean.
3. Write the translated result to $SAVE_CTX/translated.txt using the Write tool.

Other Types (X Post, Web, Document, Raw Text)

Translate inline (content is short, Step 8 reads full file anyway -- subagent adds overhead with no benefit):

1. Read $SAVE_CTX/extracted.txt
2. Invoke translate-content skill on the content
3. Write the translated result to $SAVE_CTX/translated.txt using Write tool

Step 8: Phase 2 -- Common Processing

8.1: Read Context Files

1. Read $SAVE_CTX/meta.json for type, title, author, url, etc.
2. Read $SAVE_CTX/extracted.txt -- ADAPTIVE by type:
   • X post / Web / Document / Raw text / Conversation: Read FULL file (typically <5K tokens)
   • YouTube: Read FIRST 100 LINES ONLY (title + channel from meta.json compensate; full transcript is 15K-60K tokens)

8.2: Read Tag Categories (Change #7)

Read 99-meta/vault-structure.md -- specifically the Tag Categories section. Use these categories as the source of truth for generating hierarchical tags.

8.3: Generate Metadata

Generate the following:

1. Filename: kebab-case, English, max 50 chars

   • YouTube: translate Korean title to English if needed
   • X post: derive from content theme, not author name
   • Web: extract semantic title from page title (strip site names, noise)
   • Document: derive from document title or filename (strip extension)
   • Conversation: based on user's requested topic

2. Tags: Hierarchical per vault-structure.md Tag Categories

   • Use category/subcategory format
   • YouTube: always include content/youtube tag
   • 3-5 tags total

3. Aliases: 2-3 alternative names

   • Translation (Korean <-> English)
   • Abbreviation
   • Search variation

4. Description: 1-2 sentence summary in Korean

5. Author (YouTube/X only): Include in frontmatter

   • YouTube: channel name
   • X post: author display name

8.4: Write Frontmatter (Change #5)

Write frontmatter to $SAVE_CTX/frontmatter.yml using Write tool:

---
title: "<title>"
date: <YYYY-MM-DD>
tags:
  - <tag1>
  - <tag2>
aliases:
  - "<alias1>"
  - "<alias2>"
description: "<Korean description>"
author: "<author>"
source: "<url>"
---

Note: Include trailing newline after ---. Omit author if not applicable. Omit source for Conversation type.

8.5: Assemble Final File (Change #5)

BODY="$SAVE_CTX/translated.txt"
[ -f "$BODY" ] || BODY="$SAVE_CTX/extracted.txt"
cat $SAVE_CTX/frontmatter.yml "$BODY" > "01-inbox/<filename>.md"

This avoids loading large content into the context window.

Step 9: Phase 3 -- Post-processing

9.1: Korean Review

Skip if: content is not Korean AND not translated to Korean. Always run for: Conversation saves.

Task tool:
  subagent_type: korean-reviewer
  model: sonnet
  description: Review Korean text
  prompt: |
    Review the Korean text in 01-inbox/<filename>.md.
    Read the file, skip frontmatter, review body text only.

Handle result:

• CLEAN: Notify user briefly
• HAS_SUGGESTIONS: Apply all suggestions using Edit tool, notify user

9.2: Rename Git Branch

Extract filename (without .md), apply branch name rules:

• Lowercase
• Replace spaces/underscores with hyphens
• Remove special characters
• Truncate to 30 chars max

git branch -m "doc/<branch-name>"

9.3: Offer Next Action

Conversation saves: Skip Deep-dive option.

AskUserQuestion with options:

• Deep-dive -- Analyze and compile understanding (not for Conversation)
• Delete -- Remove the saved note

Handle:

• Deep-dive: Invoke deep-dive-note skill with saved file path
• Delete: rm <saved_file_path>, confirm deletion

9.4: Cleanup (Change #4)

rm -rf "$SAVE_CTX"