Batch Save

Batch Save

Save multiple content URLs in parallel. Each URL runs the full /save workflow in its own isolated context directory. No session switching needed.

Step 1: Parse Input

Input is provided in $ARGUMENTS.

Parse flags

• --orig flag: If present, set keep_original_lang = true and strip from arguments.
• Default: keep_original_lang = false

Extract URLs

Split remaining arguments into a list of URLs (space-separated or newline-separated). Keep only items starting with http:// or https://.

• If no valid URLs: inform user and abort.

Step 2: Duplicate Check

For each URL, run:

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

If duplicates found, present all at once:

Found existing notes:

1. [url1] -> [path] ([title])
2. [url3] -> [path] ([title])

AskUserQuestion options: ["Skip duplicates, save new only", "Overwrite all duplicates", "Cancel"]

• Skip: Remove duplicate URLs from batch, continue with rest.
• Overwrite: Delete existing files, continue with full batch.
• Cancel: Abort.

If no duplicates, proceed silently.

Step 3: Preview URLs

Fetch a brief summary for each URL so the user knows what they're saving.

For each URL, detect type and fetch title:

• YouTube: Run defuddle parse "<URL>" --json 2>/dev/null | jq -r '[.title, .author] | join(" ||| ")'
• X/Twitter: Run defuddle parse "<URL>" --json 2>/dev/null | jq -r '[.author, .title] | join(": ")' (FxTwitter API fetches actual tweet content, unlike OG meta which is empty for regular posts)
• Web: Use curl -sL "<URL>" 2>/dev/null | sed -n 's/.*<title>\([^<]*\)<\/title>.*/\1/p' | head -1

Run all fetches in parallel (multiple Bash calls in a single message) since they are independent.

Present a preview table:

AskUserQuestion options: ["Save all", "Cancel"]

• Save all: Proceed to Step 4.
• Cancel: Abort.

The user can also type specific numbers to exclude (e.g., "1, 3 only" or "skip 2").

Step 4: Launch Parallel Saves

Process URLs in batches of 5 to avoid system resource saturation. If 5 or fewer URLs, launch all in a single batch.

For each batch:

1. Send all Task calls for the batch in a single message:

Task tool:
  subagent_type: general-purpose
  model: sonnet
  description: Save <content_type>: <URL>
  prompt: |
    Read .claude/skills/save/SKILL.md and execute the full save workflow for this URL:
    - URL: "<URL>"
    - keep_original_lang: <true|false>
    - Skip Step 9.1 (Korean review), Step 9.2 (branch rename), and Step 9.3 (next action offer)

    After saving, respond with EXACTLY this format:
    SAVED: <filename>.md
    TITLE: <note title>
    TYPE: <youtube|x-post|x-article|webpage|text>
    PATH: 01-inbox/<filename>.md

    If save failed, respond with:
    FAILED: <URL>
    REASON: <brief error description>

2. Wait for all agents in the batch to complete and collect results.
3. Launch next batch. Repeat until all URLs are processed.

Step 5: Collect Results

Parse each subagent's response. Build a results table:

Step 5.5: Korean Review

For each successfully saved file, launch korean-reviewer subagents in parallel (same batching as Step 4):

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

Handle results:

• CLEAN: No action needed.
• HAS_SUGGESTIONS: Apply all suggestions using Edit tool.

Step 6: Handle Failures

If any saves failed, offer retry:

AskUserQuestion options: ["Retry failed items", "Skip"]

• Retry: Re-launch failed URLs (same process as Step 3).
• Skip: Continue.

Step 7: Rename Git Branch

git branch -m "doc/batch-$(date +%m%d)"

Step 8: Present Summary

Display the final results table. Done. User decides next steps on their own.

Step 9: Cleanup

rm -rf .context/