freddyeleazar

unb-consultant

Universal NotebookLM-Based Consultant. Create and query AI experts via NotebookLM (Gemini Notebook). Activates on explicit user request or when the agent detects a domain that could benefit from an expert.

freddyeleazar 0 Updated 21h ago

Resources

6
GitHub

Install

npx skillscat add freddyeleazar/unb-consultant

Install via the SkillsCat registry.

SKILL.md

unb-consultant

CLI + MCP tool for creating and querying NotebookLM (Gemini Notebook) experts.

Modes

Mode 1: Manual (user explicitly asks)

  • "Create an expert on [topic]" → create_expert
  • "Ask the [name] expert about [question]" → ask_expert
  • "Generate the catalog for [name]" → generate_catalog
  • "Generate a skill for [name]" → generate_skill

Mode 2: Proactive suggestion (agent detects domain)

  • Detect domain keywords from user's project description
  • Call suggest_experts(domain_keywords=[...])
  • If matched, suggest to the user: "I can create a NotebookLM expert on [domain]..."
  • NEVER create an expert without explicit user confirmation.
  • NEVER write files to the project without explicit user confirmation.

Available Tools (MCP)

Tool Confirmation Required Description
create_expert Yes Create a new NotebookLM expert
ask_expert No Ask a question to an expert
list_experts No List all registered experts
add_sources Yes Add sources to an existing expert
generate_catalog Yes Generate thematic catalog
delete_expert Yes Delete an expert
generate_skill Yes if writing files Generate SKILL.md for a project
suggest_experts No Suggest creating an expert by domain
auth_check No Check authentication status

Known Domains (suggest_experts)

Keywords Suggested Name Sources
CVSS, CVE, vulnerability cvss-v4 first.org CVSS spec
FAIR, risk analysis fair-analysis fairinstitute.org
MusicXML, music notation musicxml musicxml.com
Skyrim, Creation Kit skyrim-creation-kit ck.uesp.net
0 A.D., modding 0ad-modding 0ad wiki

Common Workflow

User: "I'm working on CVSS vulnerability scoring"

Agent: (calls suggest_experts(["cvss", "vulnerability", "scoring"]))
       → matched: cvss-v4
       → "I can create a NotebookLM expert on CVSS v4.0 using
          FIRST.org's official specification. This would let you
          query the standard without loading the 70-page PDF.
          Would you like me to create it?"

User: "Yes, go ahead."

Agent: (calls create_expert("cvss-v4", urls=[...], auto=True))
       → Expert created.
       → Catalog generated.
       → Skill generated.
       → "Done. You can now ask questions like:
          'Ask the cvss-v4 expert how the base score is calculated'"

### Suggested next steps (when create_expert returns without --auto)

When `create_expert` returns `suggested_next_steps`, inform the user and ask
if they want to proceed with catalog and/or skill generation. For each step:
- Read the description aloud
- Ask "Would you like me to do that?"
- **NEVER** execute these steps without user confirmation.

User: "Ask cvss-v4 how PR is scored"

Agent: (calls ask_expert("cvss-v4", "How is PR metric scored in CVSS v4.0?"))
       → Returns answer with citations.

CLI Fallback (when MCP tools are not available)

Some models or providers do not support MCP tools. In that case, use these
CLI commands directly via the bash tool instead:

Equivalence Table

MCP Tool CLI Command
ask_expert unb ask "<name>" "<question>" [--json]
list_experts unb expert list
create_expert unb expert create "<name>" --url "..." [--file ...] [--auto]
delete_expert unb expert delete "<name>" [--yes]
generate_catalog unb catalog "<name>" [--output ./catalog.md]
generate_skill unb skill-gen "<name>" [--auto]
add_sources unb source add "<name>" --url "..." [--file ...] [--directory ...]
suggest_experts unb suggest <keyword> [<keyword> ...]
auth_check unb auth check --test --json

Rules

  1. Prefer MCP tools when available (they are faster and more structured).
  2. If MCP tools are not visible, fall back to CLI commands.
  3. Always quote arguments with double quotes in PowerShell.
  4. NEVER execute destructive commands (delete, overwrite) without user confirmation.
  5. Use --json with ask to get structured answers with source citations.

Example (no MCP)

User: "List available experts"
Agent: → unb expert list
       → Shows table of registered experts

User: "Ask the 0ad-modding-reference expert about VFS"
Agent: → unb ask "0ad-modding-reference" "How does VFS work?" --json
       → Returns answer with citations

Installation (for user reference)

pipx install "unb-consultant[browser]"
unb login

Categories

CLI Tools Code Gen
GitHub

Install

npx skillscat add freddyeleazar/unb-consultant

Install via the SkillsCat registry.

Recommended Skills