Creating a Virtual TA with Custom GPTs

A walkthrough for faculty who want to build a course-specific virtual teaching assistant using OpenAI’s Custom GPTs. The same conceptual moves apply to Claude Projects (Anthropic) or Gemini Gems (Google), but the specific UI flows differ — adapt the steps below to your platform. The example throughout is the Lab’s IP TA — a study-and-review assistant for Intro to Intellectual Property at Penn Carey Law.

This is the conceptual precursor to the Lab’s Heron virtual-TA Slackbot project. The Custom GPT path is a low-effort starting point: no code, no hosting, no integration work. You drop in syllabus materials, write a few paragraphs of instructions, and the platform handles the rest.

Before you start

Custom GPTs are useful when:

• You want students to be able to ask questions about course-specific material — your syllabus, your slides, your hypotheticals — and get answers grounded in your framing of the doctrine.
• You don’t need analytics, persistent memory across students, or integration with another platform.
• You’re comfortable with students using the AI tool’s web interface rather than having the TA embedded somewhere (Canvas, Slack, your course site).

If you need any of those things, a Custom GPT is the wrong tool — look at Heron-style implementations or build something custom.

Seven steps

1. Navigate to the GPT editor

Go to chatgpt.com/create (or open Explore GPTs from the ChatGPT sidebar and click “Create”). You need a paid ChatGPT account (Plus, Pro, Business, Enterprise, or Edu) to publish a Custom GPT. To find GPTs you’ve already built, open Explore GPTs in the ChatGPT sidebar and select My GPTs.

2. Create a new GPT

Click the plus sign to start a new GPT.

3. Open the Configure tab

Two tabs appear at the top left — Create and Configure. Skip the Create tab (it’s a guided builder that walks you through the same fields conversationally) and go directly to Configure. You’ll have more control and a cleaner result.

4. Fill in the configuration

This is where the real work happens. The fields are:

Name

Pick something descriptive. The Lab’s IP TA is called “Intro to IP TA” — students see the name in the chat header, so make it obvious what course it serves.

Description

One sentence on what the GPT is for. The IP TA’s description: “Study and Grading assistant for IP Law at Penn Carey Law.”

Instructions

This is the most important field. Treat it as a full briefing for a smart but inexperienced TA who has never seen your course before. Cover:

• Audience. Who is this for? “Students in the Intro to Intellectual Property class at the University of Pennsylvania Carey Law School.”
• Purpose. What does the TA do? Study help? Practice questions? Grading rubric explanations? Be specific. Multiple purposes are fine — just enumerate them.
• Voice and depth. How should the TA respond? “Detailed but concise explanations, enriched with real-life case examples. Encourages understanding through Q&A; helps students grasp key takeaways without overwhelming them.”
• Subject scope. What content domain? For the IP TA, the syllabus modules: patents, trademarks, copyrights, trade secrets, and basic IP-law principles.
• Boundaries. What should the TA not do? Common ones: don’t write papers, don’t answer take-home exam questions, don’t substitute for the assigned reading.

The IP TA’s full instructions live in the Lab’s internal materials — ask if you’d like to see them as a starting template.

Conversation starters

Pre-written prompts that appear as buttons when a student opens the GPT. They serve two purposes — orienting the student to what the TA can do, and seeding the conversation in productive directions. A few examples from the IP TA:

• Welcome: “Hello! I’m [Name], your virtual TA for Intro to Intellectual Property. I’m here to help you understand topics like patents, copyrights, and trademarks. Ask me about key concepts, definitions, and examples.”
• Guidance: “When asking about a specific concept, use keywords and I’ll do my best to clarify. For example: ‘What is fair use?’”
• Learning objectives: “This course covers intellectual property types, the purpose of IP rights, and the basics of protecting creations and inventions under IP law.”

Three or four starters is plenty. More clutters the interface.

Knowledge

Upload course materials — syllabus, key textbook chapters, lecture slides, practice questions, and your own hypotheticals. The GPT uses retrieval over these documents to ground answers in your course rather than the model’s general training.

Do not upload:

• Class recordings or transcripts of class recordings (FERPA: identifiable student speech is a protected education record; recording-consent terms also apply).
• Student work, exam answers, or grading materials with student names attached.
• Copyrighted casebook chapters or commercial materials beyond fair-use limits.
• Anything covered by NDA, embargo, or partner-confidentiality.

A few practical notes:

• PDFs and Word docs work well. Slides export to PDF cleanly.
• The retrieval is imperfect — if a student asks about a doctrine you covered in class but not in any uploaded document, the GPT will fall back on the model’s general knowledge, which may or may not match your framing. Upload broadly within the limits above.
• Keep an eye on file count and size. Performance degrades with very large knowledge bases.

Capabilities

Check Web Browsing, Image Generation (DALL·E), and Code Interpreter & Data Analysis. Web Browsing is the most useful — it lets the TA pull recent cases and developments. The other two are rarely used in a doctrinal-law TA but don’t hurt to enable.

Actions

Skip this. Actions are for connecting the GPT to external APIs (authentication layers, custom databases). For a study TA, you don’t need it.

5. Publish

Click Create at the top right. Choose visibility — “Only me,” “Anyone with a link,” “Everyone” (GPT Store) for personal accounts, or “Anyone at [Workspace]” for Business/Enterprise/Edu accounts. Link-based sharing is the practical default for a course TA.

6. Tune the response style

Once the GPT is live, open it and have a few real conversations. Adjust the tone in the prompting interface — “Be more concise.” “Use fewer bullet points.” “Always cite the case name when discussing doctrine.” These adjustments fold back into the GPT’s instructions automatically.

A few tone defaults that work well for a law-school TA:

• Approachable, clear, professional.
• Match the complexity students are expected to handle. Don’t dumb it down; don’t pile on jargon.
• End answers with an invitation to follow up: “If this wasn’t clear or you’d like a deeper explanation, ask me to elaborate.”

7. Test and iterate

Simulate the questions students are likely to ask. Stress-test the boundaries — try to get the TA to write a paper, or answer an exam question, or hallucinate a case. Where it fails, tighten the instructions. Where it succeeds, note what’s working.

Update knowledge as the course evolves. New cases, new hypotheticals, revised slides — re-upload and the TA stays current.

Boundaries are guidance, not enforcement. Custom GPT instruction-following on adversarial student prompts is unreliable. A determined student can often coax the TA into answering exam questions or writing a draft paper despite explicit instructions to the contrary. Treat the boundaries as design intent and student-facing norms, not enforcement; pair the TA with course-policy guidance (the AI Syllabus Guide covers academic-integrity expectations) and reserve graded assessment for tasks the TA cannot easily do.

Privacy note

OpenAI’s data-handling rules depend on plan tier. On Business, Enterprise, and Edu accounts, uploaded documents are not used for model training by default. On Plus and Pro consumer accounts, training-data use depends on your account-level Data Controls setting; opt out before uploading anything. Confirm OpenAI’s current data-handling terms for your plan tier before uploading anything sensitive — vendor policies change.

The same constraints apply to Claude Projects (Anthropic) and Gemini Gems (Google), but the defaults differ. Check the platform’s current privacy terms before uploading. Review the do-not-upload list in the Knowledge section above before adding any course material.

Status

Maintained for the Penn Carey Law community. Vendor UI flows and privacy defaults change frequently — verify against current vendor documentation before relying on the specifics here.