Best Practices for Using AI Tools

A practical guide for getting better results, faster, all while staying within usage limits.

This guide is for anyone using Penn State AI tools, including AI Studio, Copilot, and other University-approved platforms. While examples may reference specific tools, most of the practices in this guide apply to any AI chatbot.

Penn State manages access to several AI tools for University use. To help ensure these tools remain available to the entire University community, usage limits are in place. AI Studio includes monthly token limits, and other tools may have their own usage limits as well. Many people reach their limits sooner than expected, not because they are doing too much work, but because of habits that quietly increase token usage. The small changes in this guide can help you get better results, use tokens more efficiently, and make your access go further.

Quick Reference: The Five Habits

If you only remember five things, remember these:

Start fresh chats for new tasks.

Long chats slow down, get less focused, and use more of your allotment per message.
Be specific upfront.

One clear prompt beats five vague follow-ups.
Only share what's needed.

Don't upload a 200-page PDF if your question is about page 12.
Pick the right model for the job.

More advanced models typically use more of your allotment per message — match the model to the task.
Verify before you trust.

AI can sound confident and still be wrong. Always check important facts.

Best Practices:

AI tools work with the conversation context you have built. Longer chats generally consume more tokens because there is more prior context for the tool to consider. They can also become slower or less focused over time. Starting fresh when the task changes is one of the simplest ways to use your allotment more effectively.

Start a new chat for each distinct task

If you're done with one project and moving to something unrelated, open a new conversation. Don't ask the AI to help you write an email, then debug Python, then plan a meeting — all in the same thread.

DO:

Open a new chat for each distinct task or topic. When you switch from writing a memo to analyzing a spreadsheet, switch chats, too.

DON'T:

Use one chat all day for every question that pops into your head — writing, coding, summarizing PDFs, brainstorming, casual questions — until the AI starts forgetting what you said an hour ago.

Wrap up long chats before they get unwieldy

If a chat is getting long (roughly 30+ back-and-forth messages, or you notice responses slowing down), wrap it up cleanly before continuing. Two ways to do this:

Export the chat. In AI Studio, click the Export Chat icon at the top of the conversation (next to the model selector) to save a copy for your records. This is helpful if you need to revisit the work later without keeping the chat open and active.
Summarize and continue elsewhere. Ask the AI to summarize what you've covered, then paste that summary into a fresh chat to keep working.

Don't re-explain context that's already there

Within a single chat, the AI already remembers what you said. You don't need to re-paste your document, re-state your role, or re-describe the project on every message. Just ask the next question.

A specific first prompt almost always beats a vague prompt plus five clarifying rounds. Each round uses more of your allotment.

Use the Creative CRAFT framework

Penn State's AI Essentials course teaches the Creative CRAFT framework for writing strong prompts. Each element closes a specific gap that leads to vague results.

Context

Give the AI the background it needs. Context grounds the AI in your situation. Without it, the response is generic, written for anyone and useful to no one.
Role

Tell the AI who it is. A role shapes the voice and depth of the response. "Act as a data analyst" signals rigor and precision, not just a casual summary.
Action/Task

Replace vague instructions with specific ones. Vague tasks produce vague outputs. Naming exactly what to find gives the AI a concrete target, and you a useful result.
Format

Specify how the output should be structured. Named sections make the output immediately usable. Without format guidance, the AI decides the structure, and it may not match your needs.
Tone/Steps/Constraints

Set the style and limits. Without tone guidance, style is unpredictable. "Concise, professional, easy to scan" keeps the output fit for a working group (and not a textbook.)
Creative Direction

This is the final push that moves output from description to insight. This is what separates a summary from something useful. It tells the AI to go beyond reporting data and make it meaningful for the people who need to act on it.

DO:

"Draft a 200-word email from me (staff in the Provost's office) to department chairs, announcing the new AI tool policy effective May 1. Friendly but professional tone. End with a link placeholder and an invitation to office hours."

DON'T:

"Write me an email about the new policy."

One example beats a paragraph of description

If you have a sample of what good output looks like (a past email, a section of a report, a writing style you like) share it in the prompt. A single concrete example often does more than several sentences of explanation.

Ask the AI to show its approach for complex tasks

For analysis, decisions, or multi-step problems, ask the AI to outline its approach, list its assumptions, show key steps, and provide a concise rationale before the final answer. You'll get noticeably better results and a clearer view of how it got there.

Iterate, don't restart

If the first answer is close but not quite right, refine it. "Make it shorter," "more formal," "focus on X instead of Y." That's much more efficient than starting over with a new prompt.

Uploading a large PDF to ask about one section can use more resources than necessary. Large attachments may continue to affect token use, processing time, and model performance in follow-up turns, depending on the tool and model.

Trim before you upload:

DO:

Copy just the enrollment section (a few pages) into the chat, or extract the specific pages first. Same answer, far less of your allotment used.

DON'T

Upload the full 180-page annual report and ask: "What does it say about Q3 enrollment?"

Use the right format

Plain text or markdown is most efficient. If you can copy-paste instead of uploading, do it.
PDFs vary wildly. Text-based PDFs are fine. Scanned PDFs or image-heavy ones use far more resources because they have to be visually processed.
Spreadsheets can be huge. If you only need a few columns or rows, paste just those instead of uploading the whole file.
Images use more than text. A photo of a paragraph uses more of your allotment than the paragraph itself. Type it out when you can.

Don't keep large attachments in long chats

If you uploaded a document to extract something specific, get what you need, then start a new chat for follow-up work. The attached file stays part of the conversation and can keep affecting performance and token use as the chat continues.

AI Studio gives you access to multiple AI models from leading providers within one secure Penn State environment. Because model availability can change over time, this guide does not list individual models.

For the current list of available models, recommended uses, and notes about availability, see the AI Studio: Models Knowledge Base article.

The model menu at the top of AI Studio lets you switch between available models. Availability may vary depending on the AI Studio experience you are using, including ONEchat, Personal Agents, and Platform Agents.

Quick tips

Use a smaller or faster model first for everyday work such as drafting, rewriting, summarizing, brainstorming, and simple questions. These models are often more than capable for routine tasks and may use less of your monthly token allotment.
Move up to a more advanced model when the task genuinely calls for it, such as complex analysis, long documents, nuanced writing, research synthesis, advanced coding, or multi-step reasoning.
For tasks you do often, try the same prompt with more than one available model. You may find that a faster model handles most of your work well, while a more advanced model is worth saving for higher-complexity tasks.

Compare tools, not just models

AI Studio isn't your only option. Other Penn State-provided AI tools each have strengths for different workflows. For a full side-by-side comparison, see the AI Comparison Table.

: Penn State's enterprise AI tools are designed with strong data privacy controls. Your personal information is not used to train AI Studio or other AI tools. But you are still responsible for what you put into a prompt, and not all data is permitted in these tools.

The key rule

Penn State's AI tools (AI Studio, Copilot, M365 Copilot, Google Gemini) are approved for Level 1 and Level 2 data only. They are never authorized for Level 3 or Level 4 data (sensitive or restricted information such as SSNs, health records, financial information, or regulated research data). Free personal AI accounts should not be used for any Penn State work.

For the full breakdown of what each level includes and what is approved for which type of AI account, see Penn State's Data Protection page.

When in doubt, don't paste

If you're unsure whether something belongs in an AI tool, leave it out and check the Data Protection page first. Common items that should never go into any AI tool include Social Security numbers, financial account or payment card details, health information (HIPAA / PHI), driver's license or passport numbers, and passwords or API keys.

For risk assessment guidance on new AI tools or use cases, see the AI Risk Assessment Process page.

AI tools can produce confident, well-written text that is factually wrong. This is sometimes called "hallucination." It's most common with citations, statistics, names, dates, and anything specific.

Always double-check

Citations and quotes. AI tools regularly invent journal articles, court cases, and book passages that sound real but don't exist.
Numbers and statistics. Especially anything specific ("73% of students…"). Find the original source.
Names, titles, and dates. Particularly for people, organizations, and historical events.
Code that touches important systems. Test in a safe environment first.
Anything you're going to publish. Your name will be on it. Make sure it's right.

If you're using AI for coursework, research, or any Penn State publication, follow Penn State's guidelines on AI use and disclosure.

Check your instructor's or department's policy before using AI on assignments.
Policies vary widely by course. When you're unsure, ask the instructor.
When AI use is allowed, indicate which tool was used and how it was used. This is the disclosure standard described in Penn State's academic integrity guidance for students.
Disclose AI use when required by a journal, conference, grant agency, or course — and follow the citation guidance Penn State University Libraries provides for APA and MLA.
Remember that you are responsible for accuracy, originality, and appropriateness of anything you submit, regardless of how it was drafted.

For full guidance, see integrity.psu.edu/ai and the AI Guidelines page.

Common Mistakes That Use Up Your Allotment

A quick checklist of habits worth breaking:

The forever chat.

Using one conversation for weeks. Start fresh sessions for distinct tasks.
The kitchen-sink upload.

Attaching every related document "just in case." Upload only what's needed for the current question.
The five-word prompt followed by ten clarifications.

Write a fuller prompt the first time.
The screenshot of text.

If it's text, paste it as text. Screenshots use more resources and are harder for the AI to read accurately.
The "please" and "thank you" run-on.

Politeness is fine, but a separate "thanks!" message is still a full round-trip. Save it for the end.
Using the most powerful model for everything.

Reach for a smaller model for routine tasks.
Asking the AI to repeat or reformat its previous answer instead of editing the original.

Copy the response, edit it yourself, save the round-trip.