Tag Sync

Fetch tags from Raindrop, curate the top N by usage, characterize each with a one-line description, and write or sync the vocabulary file. Follows the vendor-sync pattern: registry (vocabulary file) -> fetch from external (Raindrop API) -> diff against local -> preview -> apply after approval.

The output file at ~/.claude/references/raindrop-tags.md is consumed by /session-bookmarks when selecting tags for new bookmarks.

Arguments

• /tag-sync — sync existing vocabulary (or create if no file exists)
• /tag-sync 100 — set tag count to 100, then sync/create
• /tag-sync --reset — force full recreation (re-characterize + re-cluster all)

Edge Cases

• Raindrop MCP unavailable — report error and exit. Do not write a partial file.
• find_tags returns fewer than tag_count — use all available tags. Store the actual count in frontmatter so the next sync has a realistic target.
• ai-bookmarked not in top N — always inject it regardless of rank. It is listed in mandatory_tags frontmatter.
• No bookmarks found for a tag (characterization step) — infer from tag name alone. Flag with * in preview: *auto-inferred from name.
• ~/.claude/references/ directory missing — create it before writing.
• Vocabulary fetched today, no explicit args — warn: "Vocabulary was already synced today. Pass a count or --reset to force." Do not re-fetch.
• User declines all changes — report "No changes made" and exit.

Workflow

Step 1: Parse arguments

Accept an optional integer argument or --reset flag.

• Integer sets tag_count (how many tags to include).
• --reset forces creation mode even if the file exists.
• No argument: use tag_count from existing file, or default 75.

Step 2: Load existing vocabulary

Read(file_path="~/.claude/references/raindrop-tags.md")

• File exists — enter sync mode. Extract tag_count, fetched_at, and existing tag entries (tag name, count, characterization, cluster). Also extract optional config fields from frontmatter: blocklist, context_tags, conventions. Preserve these unchanged through the sync cycle — they are user-authored config, not Raindrop-derived data. Record which of the three are absent from frontmatter — Step 8 will backfill them from the seed values defined below. Check staleness: if fetched_at matches today's date and no explicit arguments were passed, warn "Vocabulary was already synced today. Pass a count or --reset to force." and Stop. Do not proceed to Step 3.
• File missing (or --reset) — enter creation mode. Seed default config fields: blocklist: ["5", "4", "3", "2", "imported", "toread", "unread", "for:*"], context_tags: [], conventions: [].

Step 3: Fetch tags from Raindrop

mcp__raindrop__find_tags()

Returns tags with usage counts. Paginate if has_more=true. Sort by count descending. Take the top tag_count entries. Always include ai-bookmarked in the target set regardless of rank.

Step 4: Diff (sync mode only)

Compare target set against existing vocabulary:

Category	Action
New tags (in Raindrop, not in file)	Must characterize (Step 5)
Dropped tags (in file, below threshold now)	Candidate to remove
Retained tags (in both)	Keep characterization, update count
Count-only changes	Apply silently, no user action needed

If no new tags and no dropped tags (only count changes), report "Tag vocabulary is in sync — only usage counts updated" and offer to apply count updates in place.

Step 5: Characterize new tags

For each new tag (creation mode: all tags; sync mode: only new arrivals), sample 2-3 bookmarks to ground the description:

mcp__raindrop__find_bookmarks(has_tags=["<tag-name>"], limit=3, sort="random")

From bookmark titles and excerpts, write a concise one-line characterization (10-15 words max). If find_bookmarks returns empty, infer from the tag name (e.g., conways-law -> "Org structure effects on system architecture").

Do NOT call find_bookmarks for tags that already have characterizations in the existing file (sync mode).

Step 6: Assign clusters

Four groups:

Cluster	What goes here
Ecosystem	Language, framework, platform, tool names
Architecture	Design patterns, methodology, structural concepts
Content Type	Article genre, reference type, quality tier
Other	Anything that doesn't clearly fit above

• Retained tags keep their existing cluster (stability across syncs).
• New tags are auto-assigned from name + sample bookmark content.
• --reset forces re-assignment of all clusters.

Step 7: Preview

Show the proposed state before writing.

Creation mode — full vocabulary table grouped by cluster.

Sync mode — diff summary first, then full table:

## Tag Vocabulary Preview

**Mode:** Sync | **Count:** 75 | **Fetched:** 2026-04-06

### New tags to add (3)

| Tag | Count | Cluster | Characterization |
|-----|-------|---------|-----------------|
| `basis-ecosystem` | 12 | Architecture | Foundational stack choices and tradeoffs |

### Tags to remove (1)

| Tag | Count | Cluster | Characterization |
|-----|-------|---------|-----------------|
| `old-unused-tag` | 2 | Other | Below threshold |

### Count updates (silent, no approval needed)

- `javascript`: 398 -> 412
- `nodejs`: 280 -> 287

### Summary after changes

75 tags: 31 Ecosystem | 24 Architecture | 12 Content Type | 8 Other

Approve all, or specify adjustments (e.g. "keep old-unused-tag",
"move basis-ecosystem to Content Type", "change characterization for X").

Wait for user response before writing.

Step 8: Write

After approval, build the complete file and write atomically:

Write(file_path="~/.claude/references/raindrop-tags.md", content="...")

Full-file overwrite (not incremental edit). Create ~/.claude/references/ directory if it does not exist.

Reincorporate preserved config fields into the written file's YAML frontmatter. Do not overwrite or strip blocklist, context_tags, or conventions — these are user-authored and must survive the sync cycle.

Additively heal pre-seeding vocab files. If sync mode encountered a file that pre-dates the Step 2 seeding behavior, one or more of blocklist, context_tags, or conventions will be absent from the frontmatter extracted in Step 2. For each of these three fields, apply this rule when building the written frontmatter:

• Field present (even if empty — blocklist: [] or blocklist: with no value) — preserve as-is. Treat presence as an intentional user choice.
• Field absent (key not in the frontmatter at all) — backfill with the Step 2 creation-mode seed value (the same defaults applied when seeding a new file: blocklist: ["5", "4", "3", "2", "imported", "toread", "unread", "for:*"], context_tags: [], conventions: []).

Healing is additive only — never replace, merge, or augment an existing field's contents. The three fields are evaluated independently (a file may have context_tags set but be missing blocklist). Mention any healed fields in the Step 9 report so the user sees what was backfilled.

Step 9: Report

## Tag Vocabulary Synced

File: ~/.claude/references/raindrop-tags.md
Tags: 75 (31 Ecosystem | 24 Architecture | 12 Content Type | 8 Other)
Mandatory: ai-bookmarked (always applied)

Changes: added 3, removed 1, updated 18 counts

Next: `/session-bookmarks` will use this vocabulary when creating bookmarks.

Output File Format

The vocabulary file uses YAML frontmatter for machine-readable metadata and markdown tables for human-readable content:

---
tag_count: 75
fetched_at: 2026-04-06
mandatory_tags:
  - ai-bookmarked
blocklist:
  - "5"
  - "4"
  - "3"
  - "2"
  - imported
  - toread
  - unread
  - "for:*"
context_tags: []
conventions: []
---

# Raindrop Tag Vocabulary

Always include `ai-bookmarked`. Choose 2-5 additional tags per bookmark.

## Ecosystem

| Tag | Count | Characterization |
|-----|-------|-----------------|
| `javascript` | 412 | Core JS language, runtime, ecosystem |
| `nodejs` | 287 | Node.js server-side runtime and tooling |

## Architecture

| Tag | Count | Characterization |
|-----|-------|-----------------|
| `protocol-design` | 156 | Specs, RFCs, protocol architecture |

## Content Type

| Tag | Count | Characterization |
|-----|-------|-----------------|
| `foundational-text` | 89 | Seminal papers, essays, reference works |

## Other

| Tag | Count | Characterization |
|-----|-------|-----------------|
| `indieweb` | 67 | IndieWeb movement, ownership, POSSE |