Doing More with HokieAI

You have had a first chat and maybe built a first personal agent. This page is the rest of HokieAI — the parts worth learning once you have a reason to.

If you have not done that first chat, start with Getting Started. For everything that is not HokieAI — ARC, NotebookLM, Copilot, Google AI Pro — see Other AI Tools.

Chat or agent?

HokieAI’s chat interface has a name: ONEChat. It is what the New Chat button in Getting Started opens, so if you have already had one conversation you have used it. Virginia Tech’s 4Help guide calls it “the primary conversational interface within HokieAI” and says it “does not require file uploads, configuration, or prompt engineering to get started.”

Most people stay there and never need anything else.

The dividing line is repetition and reference material:

What you’re doing	Use
A quick question or explanation	ONEChat
Brainstorming or drafting something one time	ONEChat
The same task, over and over, the same way	An agent
Work that starts from files you upload	An agent
A workflow specific to your role or your department	An agent

That table is 4Help’s, and it is a good rule. The cost of an agent is the ten minutes you spend writing its instructions. If you will only ever do the task once, skip it.

Working from files

Agents can reference files you upload. 4Help’s supported file types article lists PDF, Word, PowerPoint, Excel, plain text, Markdown, CSV, JSON, XML, and HTML, plus PNG and JPEG with a caveat — and notes the list “is not complete.”

What actually determines whether this works well:

• Text beats pictures of text. 4Help’s guidance is that “PDFs with selectable text and DOCX files yield the best results,” and that you should “avoid scanned documents unless optical character recognition (OCR) has already been applied.” A scanned memo is an image, and the agent will struggle with it.
• Small and focused beats big and general. “Large or unfocused documents can reduce response accuracy.” Split unrelated topics into separate files.
• Images are accepted but weak. If the thing you need is text inside a screenshot, expect limited results.

If an agent gives vague answers about a document, the file is usually the problem, not the prompt.

Building an agent that holds up

Getting Started walks through creating one. The difference between an agent you use twice and one you use every week is what you do after it exists:

1. Write the boundaries, not just the job. Say what it should not do — don’t invent numbers, don’t assign owners that aren’t in the source, don’t exceed a length. Vague instructions produce confident guesses.
2. Test with something real and unremarkable. Use an actual past task where you already know the right answer, so you can tell whether the output is good.
3. Fix the instructions, not the output. When the answer is wrong, resist editing the result by hand. Change the instruction that let it be wrong, then run the same test again.
4. Add knowledge sources only when you need them. An agent with no reference material is easier to debug. Add files once its behavior is right.

Two limits worth knowing before you plan around them. There are two kinds of agents, and 4Help is explicit that “at this time, individuals are not authorized to create official agents” — official agents are “reviewed and managed by a department, service owner, or administrator.” And you cannot add models: HokieAI’s model selection is managed centrally, not requested per user.

What else is in there

HokieAI’s interface includes Add Files & Images, Web Search, Data & File Tools, Quick Prompts, Skills, and Reasoning Mode, along with Artifacts, Official Agents, Knowledge Sources, Add a chat starter, and Sharing. Treat these as things to reach for when a task calls for them, not a checklist. Interface labels and available options change.

Check the work before it leaves your hands

This is the part that matters most and gets skipped most. 4Help’s responsible use guidance says to “treat AI outputs as a first draft, not a final product,” to “cross-check critical information against trusted sources,” and to avoid using HokieAI for “final decision-making without human review.”

In practice: verify names, dates, numbers, quotations, and any policy claim against the real source before the thing you wrote goes anywhere. An agent that is right nine times out of ten is still going to be wrong in front of someone eventually, and it will be wrong confidently.

Data

VT has approved HokieAI for high-risk data, and states that with HokieAI, “inputs and outputs do not leave VT and are not used to train models.” That does not mean everything belongs in it. Use the smallest amount of data that will do the job, and check Virginia Tech’s current AI tools and data-risk guidance when university data is involved.

The fuller picture — which tools are approved for what, and where the lines are — is on Other AI Tools.

Questions? Open a ticket or email aadithelp@vt.edu.

Last updated: July 2026