Raindrop bookmark curator

Raindrop Bookmark Curator

Use this skill when the user wants an agent to organize Raindrop.io bookmarks, especially saved social posts, research links, AI/tool discoveries, reading lists, or inbox-style collections.

The core job is to turn messy saved bookmarks into useful library entries:

• content-based title;
• concise bookmark note/description;
• useful tags;
• appropriate collection;
• optional Markdown/Obsidian note;
• durable state so future runs process only deltas.

Public-safe constraints

• Do not commit OAuth tokens, API keys, cookies, real Raindrop export/state files, generated private notes, personal collection IDs, or private filesystem paths.
• Use placeholders such as <vault>, <state-file>, <collection-id>, and <bookmark-id> in reusable templates.
• Never delete bookmarks or collections unless the user explicitly asks and confirms the exact IDs.
• Never bypass login walls, paywalls, CAPTCHAs, private posts, or account-only comments.
• For social posts, extract only publicly visible metadata/content. Mark inaccessible comments/media as not publicly visible / not extracted.
• Do not invent post text, comments, links, engagement counts, claims, or summaries.
• Verify every mutation by reading the bookmark back from Raindrop.

Prerequisites

Preferred interface: Raindrop's Streamable HTTP MCP endpoint via mcporter:

npx -y mcporter list \
  --http-url https://api.raindrop.io/rest/v2/ai/mcp \
  --name raindrop \
  --schema --json

On first use, the tool may require OAuth in the user's browser. Do not ask the user for passwords or tokens.

Standard workflow

1. Establish scope

   • Recent/delta scan: usually find_bookmarks with has_tags, collection_ids, recency, or limit.
   • Historical reprocess: read bookmark IDs from the state file and fetch exactly those IDs.
   • Single item: fetch the provided bookmark ID or URL.

2. Load or create state

   • Use templates/raindrop-state.example.json as the shape.
   • If missing/invalid, initialize safely with empty arrays/objects.
   • Preserve existing state keys; add new metadata rather than overwriting unrelated state.

3. Fetch Raindrop data

   • Use find_bookmarks for lists and current title/tags/collection.
   • Use fetch_bookmark_content for the full bookmark payload, including note/description when available.

4. Extract public content

   • Use Raindrop metadata first.
   • Use normal web extraction for accessible articles, docs, GitHub repos, product pages, PDFs, etc.
   • For Threads/Instagram/X/TikTok-like pages, use browser extraction only when needed and only for public visible content.
   • If the page is gated, summarize only metadata that is actually visible.

5. Generate curation metadata

   • title: concise, searchable, content-based. Avoid raw URLs and social shortcodes.
   • note: 1–2 sentence bookmark description explaining what it is and why it matters.
   • tags: add helpful tags; do not remove existing user tags unless explicitly asked.
   • collection: choose the best match from the configured taxonomy. Create collections only when the user asked for sorting and the needed collection does not already exist.

6. Optional note output

   • If requested, write one Markdown note per bookmark.
   • Include source metadata, original visible text/excerpt, external links found, external link summaries, useful public comments if available, TL;DR, key ideas, why it matters, and follow-up actions.
   • For idea-capture workflows, optionally generate atomic idea cards in a separate folder. Each card should be one transferable idea, source-grounded, and include: ## The idea, ## Why it matters, ## Try this, ## Use when, ## Source evidence, and ## Related.
   • Track generated card paths in state as idea_cards[bookmark_id] = [paths...], and add an ## Idea cards generated section to the source note.

7. Mutate Raindrop

   • Use update_bookmarks for title, note/description, collection, and tags.
   • Never perform delete/merge operations as part of normal curation.

8. Verify

   • Read back with find_bookmarks for title, tags, and collection.
   • Read back with fetch_bookmark_content for note/description.
   • Report any mismatch or API limitation honestly.

9. Update state

   • Add successfully processed IDs only after all required outputs/mutations are verified.
   • Store applied title, description, collection name/id, tags, timestamps, and note path if applicable.
   • Do not mark failed or partially processed bookmarks as fully processed.

Useful mcporter calls

List tools/schema:

npx -y mcporter list \
  --http-url https://api.raindrop.io/rest/v2/ai/mcp \
  --name raindrop \
  --schema --json

Find newest bookmarks with a tag:

npx -y mcporter call \
  --http-url https://api.raindrop.io/rest/v2/ai/mcp \
  --name raindrop find_bookmarks \
  --args '{"has_tags":["ai"],"limit":50,"sort":"created_desc"}' \
  --output json

Find specific historical bookmarks by ID:

npx -y mcporter call \
  --http-url https://api.raindrop.io/rest/v2/ai/mcp \
  --name raindrop find_bookmarks \
  --args '{"bookmark_ids":[123,456],"limit":150}' \
  --output json

Fetch full bookmark content/metadata:

npx -y mcporter call \
  --http-url https://api.raindrop.io/rest/v2/ai/mcp \
  --name raindrop fetch_bookmark_content \
  --args '{"bookmark_id":123}' \
  --output json

List collections:

npx -y mcporter call \
  --http-url https://api.raindrop.io/rest/v2/ai/mcp \
  --name raindrop find_collections \
  --args '{}' \
  --output json

Create missing collections, only when sorting was requested:

npx -y mcporter call \
  --http-url https://api.raindrop.io/rest/v2/ai/mcp \
  --name raindrop create_collections \
  --args '{"create":[{"title":"AI Coding & Developer Tools","description":"AI tools and workflows for coding, repo analysis, prompting, and developer productivity.","parent_id":null}]}' \
  --output json

Update a bookmark:

npx -y mcporter call \
  --http-url https://api.raindrop.io/rest/v2/ai/mcp \
  --name raindrop update_bookmarks \
  --args '{"updates":[{"bookmark_ids":[123],"update":{"title":"Clean content-based title","note":"One or two useful sentences describing the bookmark.","collection_id":456,"add_tags":["ai-coding","prompt-engineering"]}}]}' \
  --output json

Title rules

Good titles are:

• specific enough to find later;
• short enough to scan;
• based on the content, not the platform slug;
• honest about uncertainty.

Examples:

Bad original	Better title
https://www.instagram.com/reel/...	GitReverse — reverse-engineer GitHub repos into prompts
DZsJpOtgrl	Tiny World Builder — browser voxel/isometric world editor
Post by @user	ChatGPT RPG simulator GPT collection
DYjSMxx a	Backend API system-design interview reel

Description/note rules

Bookmark notes should be concise, useful, and grounded:

Open-source browser-based tiny-world builder with terrain painting, object placement, roads, traffic, save/share/remix/export, and shared-world features. Useful as a reference for game tooling and world-building workflows.

Avoid:

• hype-only copy;
• invented capabilities;
• private comments/media that were not visible;
• long pasted excerpts better stored in a separate Markdown note.

Collection taxonomy pattern

Use a user-provided taxonomy when available. Otherwise propose a small, stable taxonomy and create collections only after the user asks for sorting.

See templates/collection-taxonomy.example.json for a starter taxonomy. Adapt names to the user's domain.

Common AI/bookmark taxonomy:

Collection	Use for
AI Game Development	game agents, simulators, world builders, RPG tools, game-dev workflows
AI Coding & Developer Tools	coding agents, repo analysis, prompt engineering, developer productivity
Software Engineering & System Design	backend, APIs, architecture, scalability, interviews
AI Creative Prompts & Assets	image/video/audio prompts, spritesheets, design assets, creative recipes
AI Research & Papers	papers, benchmarks, evals, research threads
Product & Startup Ideas	market ideas, product strategy, business references

Optional Markdown/Obsidian note format

Suggested filename:

YYYY-MM-DD - raindrop-<bookmark_id> - <slug>.md

Suggested frontmatter:

---
source: raindrop
raindrop_bookmark_id: <bookmark_id>
platform: <domain-or-platform>
link: "<url>"
saved_date: "<raindrop-created-date>"
processed_at: "<local-iso-timestamp>"
tags:
  - ai
---

Suggested sections:

# <Clean title>

## Source

- Platform/domain:
- Author/account if public:
- Raindrop bookmark ID:
- Link:
- Saved date:
- Tags:

## Original visible post text / bookmark excerpt

> Only quote text actually extracted or visible.

## External links found

- <url>

## External link summaries

### <site/repo/tool>

Grounded summary.

## Useful comments / replies

Not publicly visible / not extracted.

## TL;DR

## Key ideas

## Why it matters

## Follow-up actions

State tracking

Use the state file to avoid duplicate work and maintain an audit trail. Record applied metadata after verification, not before.

Minimum keys:

• processed_bookmark_ids
• last_checked_at
• processed_notes
• collections
• bookmark_metadata

For each bookmark, store:

• last_metadata_reprocessed_at
• applied_title
• applied_description
• applied_collection
• applied_collection_id
• applied_tags
• processed_note if a note was written

Batch and delta behavior

• For small batches, process in the current agent session.
• For large batches, split into chunks and use parallel subagents only for extraction/classification. Keep Raindrop mutations in the parent agent so writes are centralized and easier to verify.
• Treat failed extraction, failed mutation, or failed verification as a blocker for that item; leave it out of processed_bookmark_ids or mark it as partial.

Final report format

Keep final reports operational and concise:

Processed: <n>
Updated titles/descriptions: <n>
Moved collections: <n>
Collections created: <n>
Notes written: <n>
State file: <path>
Verification: read back <n>/<n> bookmarks successfully
Blockers: <none or list>

When reporting public/shared output, do not include private file paths unless the user is operating locally and needs them.