This is the second in our hands-on workshop series designed to help you build practical AI tools for your actual work. Today we’re focusing on creating low-stakes assistants that can answer common questions using your materials as the source of truth.
Time and Location:
- When: Today, January 8th
- Time: 10:00 am – 11:15 am
- Where: Law Library 300 & Remote Option via Zoom
- Who: Everyone
What to expect in this session:
- Warm-up activity: Use AI to clarify what kind of resource guide assistant would be most useful for you
- Live demo: Watch the full workflow—from conversation to working assistant
- Hands-on building: Create your own prototype using Gemini Gems or NotebookLM
- Show & tell: Share what you built and troubleshoot together
- Leave with: A working prototype (or a clear plan to finish one)
What is a resource guide assistant?
Think of it as a triage-style assistant that helps answer common questions based on materials you provide. Examples include:
- A syllabus-based assistant that answers “What’s the attendance policy?” or “When is the paper due?”
- A department workflow guide that explains “How do I submit a reimbursement?”
- A course materials assistant that can explain concepts from your readings or lectures
These assistants handle the routine questions, freeing you up for the complex conversations that need a human touch.
What to bring:
A laptop, tablet, or mobile device for hands-on building. If you have specific materials in mind (a syllabus, workflow document, course readings), feel free to bring them—but it’s not required. Even a single document is enough to create a useful prototype.
No prior AI experience is necessary. This session is designed for all comfort levels and emphasizes learning through doing.
Workshop Resources
Tools Used
Detailed Summary of Topics Covered
Note: This summary was created by the “transcript-processing” skill created by Kyle Fidalgo and used with Claude Code. Content was reviewed for accuracy by a human.
Workshop Overview: “Building a Resource Guide with AI”
This hands-on workshop taught faculty how to create custom AI assistants that serve as resource guides – low-stakes tools that help answer common questions based on materials you provide. The session covered both Google Gemini Gems and Notebook LM as implementation platforms.
Core Conceptual Frameworks
The Five Design Basics for Working with Generative AI
- Sentient Design Material: AI is a probabilistic, highly adaptable tool that’s context-aware, multimodal, and flexible for knowledge work
- Natural Language Interface: If you can write clearly or think aloud into a text input, you can collaborate with AI – no technical expertise required
- Iteration as Real Work: Working with AI is simple but not always easy – planning and refinement is where the value comes from
- Meta-Prompting: When stuck, ask AI for help – it serves as both the vehicle and the guide to get you where you need to go
- Context is Everything: AI is lost without your goals, audience, and intent – you define what good looks like
The Four Ds of AI Fluency (Anthropic Framework)
- Delegation – Knowing what to offload to AI vs. keep for yourself vs. co-create together
- Description – Clear prompting and context engineering (good communication = good prompting)
- Discernment – Critical evaluation of outputs for accuracy, quality, and relevance (non-negotiable for meaningful work)
- Diligence – Taking ownership and responsibility for AI-generated output, including proper attribution
What Resource Guides Are (and Aren’t)
Good fit for:
- Answering questions about materials you provide (syllabus, policies, procedures)
- Navigating logistical aspects (office hours, due dates, submission processes)
- Explaining concepts from your content
- Quick reference for processes and procedures
- Onboarding students or colleagues
Not suitable for:
- High-stakes academic judgment or grading
- Situations requiring significant nuance beyond the provided materials
- Scenarios where a confidently wrong answer could cause harm
- Tasks that truly require human judgment (advising, admissions decisions)
Mental Models for Building AI Assistants
The Golden Rule of Prompting
If you hand off your instructions to a grad assistant or new colleague without additional context, could they complete the task successfully? If there’s room for a human to be confused, the AI will be too.
Brilliant Employee with Amnesia
Think of AI as onboarding a brilliant but very new employee who needs clear guidance. Every new chat wipes the slate clean. They’re “inexperienced” only in that they lack context about your specific job and audience – that’s your job to provide.
Degrees of Freedom (Risk-Based Design)
- High-risk/high-stakes = Give low freedom (narrow, prescriptive instructions, explicit human handoff points)
- Low-risk/low-stakes = Allow high freedom (more exploratory, trust model capabilities)
- Most useful assistants fall somewhere in the middle
Partnership Mindset
You bring: Vision, judgment, context, taste, and your expertise
AI brings: Broad knowledge, context processing ability, adaptability, 24/7 availability
Neither is complete without the other. Your resource guide needs your expertise baked into it – this is where you bring value to the AI system.
The Gem Creation Workflow
- Have a vision or idea – Or just start with the seed of something and talk it out
- Build rich context through conversation – What does the assistant need to know? About you, your audience, the scope, the materials?
- Ask AI to generate the system prompt – Use the same conversation window; AI can prompt itself well
- Set up in your preferred tool – Gemini Gem or Notebook LM
- Test and document – What works well? What doesn’t? What needs improvement?
- Feed feedback back and iterate – The description-discernment loop; refine until satisfied
Setting Up a Gemini Gem (Step-by-Step)
- Navigate to Gems – In Gemini, expand the left menu and click on “Gems”
- Create New Gem – Click the “New Gem” button in the Gem Manager
- Name your assistant – Give it a descriptive name (and optional description)
- Paste system instructions – Copy your system prompt into the Instructions field
- Tip: Ask AI to format in Markdown for better visual hierarchy
- Add knowledge sources – Click “Add from Drive” to attach up to 10 files
- Save – Click Save in the top right
- Share – Use the Share button to set access (BC-wide link or specific emails)
- Test – Start a chat and ask questions your users would ask
Setting Up a Notebook LM Resource (Step-by-Step)
- Go to notebooklm.google.com – Click “Create” to start a new notebook
- Add your sources – This is where Notebook LM shines:
- Upload files: PDFs, Word docs, text files, audio files
- Link to websites: Paste URLs for public web pages
- Connect Google Drive: Pull in documents directly
- YouTube videos: Paste video URLs for automatic transcription
- Paste text directly: Copy/paste content right into the tool
- Web search: Let it research a topic and pull sources automatically
- Understand the three-panel interface:
- Sources panel (left): Manage which sources are active in the conversation
- Chat panel (center): Ask questions grounded in your sources
- Studio panel (right): Generate additional resources
- Use the Studio to create resources:
- Audio Overview: Generate a podcast-style discussion of your materials
- Briefing Doc: Create a summary document
- Study Guide: Generate study materials from your content
- Slides: Generate a presentation
- And more…
- Share with others – Click Share and add people by email (currently requires direct invitations, not link-based sharing)
Key Differences: Gemini Gems vs. Notebook LM
| Feature | Gemini Gems | Notebook LM |
|---|---|---|
| Custom instructions/system prompt | Yes | No |
| Maximum sources | 10 files | 50 sources |
| Source types | Google Drive files | Files, websites, YouTube, web search, pasted text |
| Hallucination risk | Moderate (can draw on general knowledge) | Lower (strictly tied to sources) |
| Citation behavior | Links to source | Inline citations with exact locations |
| Sharing | Link-based (BC-wide or public) | Direct email invitations only |
| Generate additional resources | No | Yes (podcasts, slides, study guides, etc.) |
| Best for | Personalized assistants with specific behaviors | Deep research and source-grounded Q&A |
Quick decision guide:
- Need custom personality/instructions? → Gemini Gem
- Have lots of documents (10+)? → Notebook LM
- Want to create podcasts or study guides? → Notebook LM
- Need easy link sharing? → Gemini Gem
- Want strictest source grounding? → Notebook LM
Promises and Pitfalls
Promises:
- Empowers and amplifies your abilities
- 24/7 availability for common questions
- Consistent guidance based on your materials
- Frees you up for higher-value interactions
Pitfalls to design around:
- Knowledge cutoffs and training constraints
- Sycophancy – AI eager to please at expense of accuracy
- Hallucination – confident but wrong answers
- No replacement for human judgment on complex issues
Quick Tips for Putting This Into Practice
Getting Started
- Pick one real use case – A syllabus assistant, workflow guide, FAQ bot, or course materials companion
- Talk it out first – Use AI to brainstorm and clarify your vision before building
- Start small – Better to have a focused assistant that works well than an ambitious one that doesn’t
Building Your System Prompt
- Use AI to help write it – Have a context-rich conversation, then ask AI to generate the system prompt
- Ask for Markdown formatting – Visual hierarchy helps AI process instructions better
- Include human handoff points – Contact information and escalation paths so AI doesn’t have to guess
Setting Appropriate Boundaries
- Match degrees of freedom to risk – More restrictive for high-stakes, more open for exploratory
- Be explicit about what it shouldn’t do – Especially important for student-facing assistants
- Consider pedagogical approach – Do you want it to give direct answers or nudge toward discovery?
Testing and Iteration
- Test before sharing – Ask the questions your users will ask; see if answers are appropriate
- Document what works and what doesn’t – Feed this back into prompt refinement
- Don’t aim for perfection – “Good enough” gets you feedback; perfect never ships
Sharing with Others
- Create starter prompts – Help users overcome blank-input paralysis
- Explain the tool’s purpose and limitations – Set appropriate expectations
- Ask for feedback – Learning what users actually need improves future iterations
Copyright and Privacy Considerations
- Publicly available materials are generally safe – Law reviews on public websites, etc.
- BC Gemini doesn’t train on your data – Safer environment, but fair use still applies
- Be cautious with database-accessed materials – More consideration needed
- Communicate expectations to students – About not dumping copyrighted texts
Tool Recommendations
BC-Supported Tools
- Google Gemini (preferred): Use Gemini 3 Pro for serious work, Flash for quick answers
- Notebook LM: For large document sets and strict source-grounded Q&A
Model Selection in Gemini
- Pro: Best capabilities, use for reasoning-intensive work and custom assistant creation
- Fast: Quicker, for simple questions that don’t need deep reasoning
- Thinking: Flash model with extra reasoning, good middle ground
Other Tools (with appropriate caution)
- Claude: Kyle’s personal favorite, requires gettech request for BC work
- ChatGPT: Also viable, same guidelines apply
- Check bc.edu/genai for data security guidelines
Emerging Tools (Not Yet Available at BC)
- PlayLab: Shows anonymized chat logs for custom assistants (free for educators)
- Boodle Box: Multi-model access with usage insights