langtalk

langtalk — Claude Code Skill

A hybrid research skill for Claude Code combining a persistent NotebookLM notebook (seeded with the Langtalks podcast corpus) and live web search. It delivers synthesized answers on LLM engineering with inline citations ([LT*], [W*]).

Best for architecture planning, trade-offs, and lessons learned—where named-practitioner provenance complements broad web context.

---

Installation & Setup

Install the plugin via the Claude Code marketplace:

/plugin marketplace add yehuda-yu/langtalk-claude-skill
/plugin install langtalk@langtalk-claude-skill

One-Time Setup & Prerequisites:

1. NotebookLM CLI: Install the nlm CLI via pipx/uvx and run nlm login in your terminal to authenticate.

2. YouTube API Key: Obtain a v3 key from the Google Cloud Console.

3. Configure Credentials: Save the key to the plugin's persistent data directory:

   echo "YOUTUBE_API_KEY=YOUR_KEY_HERE" > ~/.claude/plugins/data/langtalk@langtalk-claude-skill/.env

4. Bootstrap the Notebook: Run the first-time setup inside Claude Code (takes ~5 minutes, idempotent):

   /langtalk:consult bootstrap

To update the plugin later, run: /plugin update langtalk@langtalk-claude-skill.

---

Usage

Ask architecture or engineering questions directly in Claude Code:

/langtalk:consult "I'm building an embedding-based ranking system. Should I embed full paragraphs or extract entities?"

Or use natural language (Claude Code will auto-discover the skill):

Use langtalk to explain reranking and help me plan the architecture.

Commands & Modes

Command	Description
/langtalk:consult "<question>"	Default: Hybrid mode (Web + Langtalks synthesis).
/langtalk:consult skill-only "<question>"	Langtalks context only (no web search).
/langtalk:consult web-only "<question>"	Web context only (skips the podcast).
/langtalk:consult update	Manually pull new Langtalks episodes.
/langtalk:consult status	Show notebook ID, source count, and last refresh date.

Note: The skill auto-refreshes silently if the last update was more than 7 days ago.

---

How it Works

When triggered, Claude Code orchestrates the process by running a bash script to query the Langtalks NotebookLM and triggering a parallel Web Search tool. It then synthesizes both data streams into a 500-700 word answer.

The output includes exact inline source tags ([LT*] / [W*]) and a unified footer with clickable YouTube URLs, showing exactly which source contributed to each claim.

---

When to Use

✅ Good fit:

• Architecture-planning ("I'm building X, what should I consider?")
• Trade-off evaluations ("LangGraph vs. CrewAI")
• Production lessons learned

❌ Skip it for:

• Quick factual lookups (standard web search is faster)
• Topics outside LLM/AI engineering
• Code completion or syntax help

---

Security

API keys and local state are stored securely in the plugin's persistent data directory (~/.claude/plugins/data/...), separate from the read-only plugin code cache. state.json and .env are gitignored and never committed. The skill never writes secrets to logs.