Session Bookmarks

Session Bookmarks

Scan the current conversation for high-signal URLs discovered during work, suggest 1-3 as Raindrop bookmarks, and create them after user approval.

This skill can be invoked directly or is auto-delegated from /session-reflect after its Basic Memory workflow completes.

Edge Cases

• No URLs in conversation — exit silently, no output.
• All candidate URLs already bookmarked — exit silently.
• Raindrop MCP unavailable — report "Raindrop unavailable — skipping bookmark suggestions" and exit. Do not block the calling skill.
• create_bookmarks succeeds but update_bookmarks (tags) fails — report partial success: "Bookmark created but tagging failed — add tags manually in Raindrop."
• User declines all — report "No bookmarks created" and exit.
• More than 3 candidates pass the filter — take the top 3 by priority: official docs > specs/RFCs > blog posts/tutorials > other.

Signal Detection

A URL is bookmark-worthy when ALL of these criteria are met:

1. Discovered during the session — via Tavily search, DeepWiki, web fetch, or Claude's own research. NOT a URL the user pasted as the starting point of the conversation. URLs found by following a user-pasted link ARE eligible (the discovery chain matters, not the seed).

2. Central to a decision or insight — the URL's content was actively used to inform a decision, solve a problem, or confirm an approach. Not just mentioned in passing or returned in a search result that was skipped.

3. Durable, authoritative content — official documentation, specifications, RFCs, well-written blog posts, research papers, detailed tutorials.

4. NOT any of these (always skip):

   • Package registry index pages (npmjs.com/package/X, crates.io/crates/X, pypi.org/project/X — these are canonical and already discoverable)
   • GitHub repository root pages (github.com/owner/repo without further path)
   • Localhost or development URLs
   • MCP tool-internal or generated URLs

   GitHub sub-pages ARE eligible: issues, discussions, release pages, specific file paths, wiki pages, PR threads.

Cap at 3 candidates maximum.

Workflow

Step 1: Extract URL candidates

Scan the conversation for URLs that meet the signal detection criteria above. Focus on URLs that appeared in tool results (Tavily, DeepWiki, web fetch) and were subsequently discussed or used.

If no candidates are found:

• Delegated from /session-reflect — exit silently (no output).
• Standalone invocation — report "No bookmark-worthy URLs found in this session" and exit.

Step 2: Deduplicate against Raindrop

For each candidate URL, check if it already exists using a two-pass approach:

Pass 1 — Exact URL match (primary):

mcp__raindrop__find_bookmarks(url_pattern=["*<url-without-protocol>*"])

Strip protocol (https://, http://) and #fragment. Keep www if present. Wrap in leading and trailing * wildcards.

Example: https://fastify.dev/docs/Reply/#send becomes url_pattern=["*fastify.dev/docs/Reply/*"].

For github.com URLs, include owner/repo in the pattern to avoid matching thousands of GitHub bookmarks: ["*github.com/fastify/fastify*"].

If any result returned, the URL is already bookmarked — skip this candidate.

Pass 2 — Domain search (fallback, only if Pass 1 returned 0):

mcp__raindrop__find_bookmarks(search="<domain-name>")

Search by domain name. Scan results for same content at a different URL (mirrors, CDN republication). A domain match alone is not a duplicate — only skip if the content is genuinely the same resource.

If the exact URL exists in ANY Raindrop collection (not just AI-bookmarked), skip it — do not create a duplicate.

Collection discovery: Before the first create_bookmarks call, discover the AI-bookmarked collection:

mcp__raindrop__find_collections(search="AI-bookmarked")

If the collection does not exist, create it lazily:

mcp__raindrop__create_collections(collections=[{title: "AI-bookmarked"}])

Cache the collection ID for use in Step 4a.

If all candidates are already bookmarked, exit silently.

Step 3: Preview

Show surviving candidates grouped in a clear format:

## Suggested Bookmarks

1. **Fastify Reply Lifecycle** — https://fastify.dev/docs/latest/Reference/Reply/
   Tags: `ai-bookmarked`, `fastify`, `reply-lifecycle`
   _Resolved the double-send issue — documents execution flow after send()_

2. **RFC 9421: HTTP Message Signatures** — https://www.rfc-editor.org/rfc/rfc9421
   Tags: `ai-bookmarked`, `protocol-design`, `foundational-text`
   _Key reference for the webhook HMAC signing design_

Approve all, or specify numbers (e.g. "1").

Each candidate shows:

• Title — auto-derived from conversation context or the URL itself
• URL — the full URL to bookmark
• Tags — always ai-bookmarked + 2-4 domain tags
• Rationale — one line explaining why this URL mattered in the session

Tag selection — hybrid approach:

Step A — Copy-from-similar (primary signal): Search for bookmarks similar to the candidate's topic:

mcp__raindrop__find_bookmarks(search="<topic keywords from title>", limit=5)

Extract tags from the top 3-5 results. Count tag frequency across results. Skip results with broken links if visible in response metadata — deprioritize tags from dead bookmarks.

Step B — Topic-match boost: If a tag name appears as a word in the bookmark's title or topic keywords, boost it regardless of frequency. This prevents specific tags (e.g., micropub) from losing to generic ones (code) that are more common across results.

Step C — Filter blocklist: Remove any tag listed in the blocklist frontmatter field of the vocabulary file (~/.claude/references/raindrop-tags.md, managed by /tag-sync — see "Integration" below). Support exact matches (case-insensitive) and prefix wildcards (entries ending in *). If the vocabulary file exists but blocklist is absent, skip filtering. If the vocabulary file is missing entirely, apply a minimal fallback: numeric-only tags (1-9) and imported.

Step D — Fallback: If fewer than 3 similar bookmarks found in Step A, match topics against the vocabulary file's table characterizations. This covers novel topics with no exemplars in the library.

Step E — Selection rules:

• ai-bookmarked is mandatory (always first tag)
• Pick 2-3 additional tags (4 more at most, beyond the mandatory tag)
• Each tag must add DISTINCT information — if A implies B, drop B
• Prefer the MOST SPECIFIC tag: nodejs not javascript
• Two tags from the same cluster is a smell — prefer cross-cluster

Step F — Examples: BAD: ai-bookmarked, tools, code, javascript (all generic, same cluster) GOOD: ai-bookmarked, nodejs, testing (specific ecosystem + concern) BAD: ai-bookmarked, mac, osx (redundant Apple tags) GOOD: ai-bookmarked, osx, cli (platform + what kind of content)

Wait for user response before proceeding.

Step 4: Create approved bookmarks

For each approved bookmark:

Step 4a — Create the bookmark:

mcp__raindrop__create_bookmarks(create=[{
  link: "<url>",
  title: "<title>",
  collection_id: <ai-bookmarked-id>,
  note: "<rationale from preview>"
}])

CRITICAL: Use the AI-bookmarked collection ID discovered in Step 2. NEVER write to any other collection — the user's organically curated library must not be contaminated with AI-generated additions.

Multiple approved bookmarks can be batched in a single create_bookmarks call (up to 50 per call).

Step 4b — Add tags:

mcp__raindrop__update_bookmarks(updates=[{
  bookmark_ids: [<returned-id>],
  update: { add_tags: ["ai-bookmarked", "<tag1>", "<tag2>"] }
}])

Always include ai-bookmarked in the tag list. Batch up to 100 tag updates per call.

Step 5: Report

## Bookmarks Created

Added N bookmarks to AI-bookmarked:
- Fastify Reply Lifecycle — <raindrop-url>
- RFC 9421: HTTP Message Signatures — <raindrop-url>

Skipped: [anything the user declined]

Use Raindrop URLs (not original URLs) in the report — per the Raindrop MCP server convention, always link to the Raindrop bookmark page.

Guidelines

• Preview before writing — never create bookmarks without user approval.
• Be selective — 1-3 high-quality bookmarks beat a dump of every URL seen.
• Rationale matters — the note field on each bookmark explains why it was bookmarked, which helps when reviewing AI-bookmarked items later.
• Tag consistently — use the vocabulary file when available. Match existing tags rather than inventing new ones.
• Graceful degradation — if Raindrop is unavailable or the tag vocabulary file is missing, the skill should degrade gracefully, not fail loudly.

Integration

• From /session-reflect — session-reflect auto-delegates to this skill in its Step 6 (Report) after completing Basic Memory work. The delegation is seamless — the user sees bookmark suggestions as a natural continuation.
• Standalone — can be invoked directly as /session-bookmarks at any point during a session.
• Tag vocabulary — consumes ~/.claude/references/raindrop-tags.md (managed by /tag-sync). Falls back to ad-hoc tags if the file is missing.