`/tips-bookmarks`¶

Save and retrieve tips / bookmarks

Pack: claudeblattman (Chris Blattman)

Category: infra

Field: general

License: MIT

Updated: 2026-04

Stages: —

↗ view SKILL.md on source · GitHub stars

Tips Bookmarks¶

v1.0 — Initial build.

Pull X/Twitter bookmarks via twitter-cli and feed new ones into a tips pipeline log. Replaces the manual copy-paste-email loop for surfacing AI/dev tips you've bookmarked but haven't acted on.

Designed to run independently from a Gmail-based curate skill (e.g., /tips-curate) — a Twitter auth failure never blocks email-based tips processing, and vice versa.

CRITICAL: No Permission Prompts¶

Do NOT use Task agents or ToolSearch. All required tools are pre-approved. Call them directly:

Bash (for twitter bookmarks --json)
Read, Edit, Write (for the state file and the tips log)

Arguments¶

$ARGUMENTS can include:

dryrun — Show what would be processed without saving or updating state
limit:N — Max new bookmarks to process (default: 10)
init — First-run initialization: mark all current bookmarks as seen, process only the 10 most recent

Instructions¶

Step 0: Pre-Checks¶

Test twitter-cli auth:
Bash
```
twitter bookmarks --json 2>/dev/null | python3 -c "import sys,json; d=json.load(sys.stdin); print(f'OK: {len(d[\"data\"])} bookmarks')" 2>&1
```
If this fails or prints an error: report TWITTER AUTH EXPIRED — refresh cookies (see your twitter-cli setup notes). and stop. Do NOT write to log or update state.
Read state file (~/.claude-assistant/state/twitter-bookmarks-seen.json):
If it doesn't exist and init was NOT given: report "No state file found. Run /tips-bookmarks init for first-time setup." and stop.
If it doesn't exist and init WAS given: proceed to Step 1 (init mode).
If it exists: load seen_ids list.
Read tips log for dedup: Read ~/.claude-assistant/tips/collected-tips-log.md and note existing URLs (to avoid duplicates from both Twitter and email paths).

Step 1: Fetch Bookmarks¶

Bash

twitter bookmarks --json 2>/dev/null

Parse the JSON. The structure is:

JSON

{
  "ok": true,
  "data": [
    {
      "id": "...",
      "text": "...",
      "author": {"name": "...", "screenName": "..."},
      "createdAt": "Fri Mar 06 18:33:43 +0000 2026",
      "urls": ["..."],
      "articleTitle": "...",
      "quotedTweet": {"text": "...", "author": {...}}
    }
  ]
}

If init mode: - Write ALL bookmark IDs to seen_ids in state file - Process only the 10 most recent (or limit:N) through Steps 2–3 - Report: "Initialized state with [N] existing bookmarks. Processing [M] most recent."

Normal mode: - Filter out any bookmark whose id is already in seen_ids - If zero new bookmarks: report "No new bookmarks since last run ([last_run date])." and stop. - Take up to limit:N (default 10) new bookmarks for processing.

Step 2: Assess & Classify¶

For each new bookmark, apply the same classification taxonomy as a sibling Gmail-based curate skill:

Classification	Criteria
`worth-reviewing`	Claude Code tip, AI workflow technique, useful dev/productivity pattern, prompting strategy
`not-relevant`	Personal, social, general news, memes, not related to AI/dev workflows
`needs-manual`	Can't determine (video/audio only with no text, link-only tweet with no context)

Quality filter (for worth-reviewing items):

High — Novel + actionable + backed by real usage
Medium — Interesting technique but may not apply directly
Low — Already known, superficial, or too vague (treat as not-relevant)

For High/Medium items, generate: - Insight (1–3 sentences): What's the technique? Be specific. - Action (1 sentence): What should you change, try, or compare? - Tags: 1–3 from: [skill-design] [prompting] [agent-pattern] [tool] [workflow] [mcp] [context-management] [coding] [productivity]

Enrichment: If a bookmark has non-X URLs in its urls array, attempt WebFetch for additional context. If WebFetch fails, classify based on tweet text alone.

Step 3: Present Recommendations¶

Text Only

────────────────────
BOOKMARKS REVIEW ([N] new of [M] total)
────────────────────

RECOMMENDED:

  1. @handle — [tweet snippet] (High)
     Insight: [what's useful]
     Action: [what to try]

  2. @handle — [tweet snippet] (Medium)
     Insight: [insight]
     Action: [suggestion]

SKIP (not relevant):
  3. @handle — [tweet snippet] — [reason]

NEEDS YOUR HELP:
  4. @handle — [tweet snippet] — video/link only, no text context
────────────────────
Which to save? (e.g., "save 1,2" / "save all recommended" / "skip all")

Wait for the user's response before proceeding.

Step 4: Save Approved Items¶

Based on the user's response:

Append approved items to ~/.claude-assistant/tips/collected-tips-log.md:

Markdown

### [Descriptive Title] [tag1] [tag2] [quality]
- **Source:** @screenName on X ([date])
- **Insight:** [1-3 sentence description]
- **Action:** [Concrete next step]
- **URL:** https://x.com/screenName/status/[tweet_id]

Group under today's date heading. Check if today's date heading already exists.

Update state file — add ALL processed bookmark IDs (saved, skipped, and not-relevant) to seen_ids. Update last_run timestamp.
If dryrun: Skip saving and state update. Just show the report.

Report:

Text Only

Saved [N] tips to collected-tips-log.md.
State updated: [total seen_ids] bookmarks tracked.
[K] remaining in current bookmark fetch not yet processed.

Step 5: Log Performance¶

Bash

echo "$(date +%Y-%m-%d),tips-bookmarks,TOOL_CALLS,NOTES" >> ~/.claude-assistant/logs/skill-performance.csv

Replace TOOL_CALLS with exact count. Replace NOTES with volume info (e.g., "50-fetched-8-new-3-saved").

State File Schema¶

~/.claude-assistant/state/twitter-bookmarks-seen.json:

JSON

{
  "schema_version": 1,
  "last_run": "2026-03-14T12:00:00Z",
  "seen_ids": ["id1", "id2", "..."]
}

Prune seen_ids older than 90 days only if the list exceeds 1000 entries (bookmark IDs are small; no urgency to prune).

Error Handling¶

Condition	Behavior
Auth failure (twitter-cli errors)	Print auth-expired message, stop. Do NOT update state.
State file missing (no `init`)	Prompt to run with `init` argument
Tips log missing	Create with standard header, continue
Zero new bookmarks	Report "no new bookmarks" with last_run date, stop
WebFetch fails on URL	Classify from tweet text alone

Examples¶

Text Only

/tips-bookmarks init        # First run: initialize state, process 10 most recent
/tips-bookmarks             # Normal run: process new bookmarks since last run
/tips-bookmarks dryrun      # Preview without saving
/tips-bookmarks limit:5     # Process at most 5 new bookmarks

Customization Points¶

State file location: Default ~/.claude-assistant/state/twitter-bookmarks-seen.json. Adjust to wherever you keep small per-skill state files.
Tips log location: Default ~/.claude-assistant/tips/collected-tips-log.md. Match to whatever path your /tips-curate (or equivalent) uses, so both Twitter and email paths feed the same log.
Classification taxonomy: The worth-reviewing / not-relevant / needs-manual split and the High/Medium/Low quality filter are designed to mirror a sibling email-based curate skill. Edit the criteria table to match your own tip-quality bar.
Tag vocabulary: Defined inline in Step 2. Add or remove tags so they line up with your tips log's existing tag conventions.
Default limit: 10 new bookmarks per run. Lower it (e.g., 5) for noisier feeds; raise it for sparser ones.
twitter-cli binary: Assumes a twitter CLI on $PATH that accepts bookmarks --json. Substitute any equivalent CLI that emits the same JSON shape (id, text, author, urls, articleTitle, quotedTweet).

/tips-bookmarks¶

Tips Bookmarks¶

CRITICAL: No Permission Prompts¶

Arguments¶

Instructions¶

Step 0: Pre-Checks¶

Step 1: Fetch Bookmarks¶

Step 2: Assess & Classify¶

Step 3: Present Recommendations¶

Step 4: Save Approved Items¶

Step 5: Log Performance¶

State File Schema¶

Error Handling¶

Examples¶

Customization Points¶

`/tips-bookmarks`¶