Week 4: Prompt Engineering
AI literacy essentials for academic libraries
Choice and Clarivate have teamed up to develop an eight-week newsletter-based course on generative AI literacy for academic library workers. If you are AI curious—and short on time—then this course is for you!
Each week contains three sections:
- Introduction: An overview of a core AI literacy competency
- In Practice: An interview or case study
- Reading Room: Further reading material focused on the topic
- The module closes with a quiz and discussion questions.
Make sure you’re registered to receive all eight weeks directly in your inbox! Know someone who might be interested in this micro-course? Share the registration page with your colleagues.
In week four, we look at best practices for prompt engineering.
Table of Contents
Introduction
Leveraging Generative AI in Libraries: Advanced Prompting Techniques and Practical Guidance
by Leo S. Lo, Dean and Professor at the College of University Libraries and Learning Sciences at the University of New Mexico, and ACRL Past-President
Generative AI tools have rapidly evolved from curiosities to useful resources in academic and public library contexts. However, realizing the full potential of these tools hinges on one key skill: prompt engineering. Effective prompts are not merely instructions, they are carefully structured communications that guide AI models to produce reliable, precise, and relevant results.
Why Prompt Engineering Is Essential
Prompt engineering involves systematically crafting and refining inputs for AI to obtain optimal outputs. Unlike human collaborators, AI models cannot intuit context or intention. Instead, they rely entirely on explicit and detailed instructions. Recent research underscores this point: ambiguous prompts yield vague outputs, while structured, detailed prompts consistently produce targeted, useful results.
Two widely adopted frameworks help ensure prompts are effective:
- CLEAR (Concise, Logical, Explicit, Adaptive, Reflective) guides users to craft concise and structured prompts, explicitly state requirements, iteratively refine instructions based on feedback, and reflect on results for future improvement.
- PTCF (Persona, Task, Context, Format) ensures that critical prompt elements—defining the AI’s persona, clearly stating the task, providing necessary context, and specifying output format—are consistently included, significantly enhancing result accuracy and usability.
Integrating CLEAR and PTCF ensures prompts are both comprehensive and precisely phrased, minimizing revisions and maximizing reliability.
Effective Prompting Strategies by AI Modality
Text-Based AI Assistants
Crafting effective prompts for text-generation AI requires explicit and iterative instructions. Rather than a vague request such as “Explain data management,” a structured prompt would be:
“You are an instruction librarian. Draft a concise (200-word maximum) introduction to data management plans targeted at biology graduate students unfamiliar with the topic. Provide three clear benefits and a short self-assessment quiz.”
This explicit approach: detailing persona, task, context, and format guides the AI toward highly targeted outputs.
An iterative approach is vital: users can first prompt the AI to outline content before requesting detailed paragraphs. Prompting the model to provide citations with verifiable DOIs and cross-checking these citations against authoritative databases prevents fabricated references, ensuring content reliability.
Image Generation Tools
When prompting AI for images, specificity transforms outcomes. Effective image prompts resemble precise scene descriptions. Rather than a generic request like “students studying,” which might give you any random image of students studying (Image 1), the user could specify:
“An illustration of two college students quietly studying at wooden desks in a cozy, warmly lit library reading room, depicted in soft watercolor style reminiscent of vintage educational posters.” (Image 2)
Explicitly mentioning artistic styles (e.g., “in the style of Edward Hopper”), lighting conditions (“soft morning sunlight”), mood (“calm and scholarly”), and perspectives (“close-up view from slightly above”) significantly improves image accuracy and visual appeal, aligning closely with desired outcomes.
Image 1
Image 2
Code Generation Assistants
Prompting code-generating AI demands clarity on the desired functionality, technical context, and specific constraints. An effective code prompt might be:
“In Python 3.12, using pandas, write a function to sort a dataset of book records by publication date. Provide clearly commented example code showing expected inputs and outputs. Avoid deprecated methods and ensure the solution handles missing publication dates gracefully.”
Incremental prompting, beginning with basic functionality, verifying accuracy, then adding features like error handling or logging, allows users to quickly identify and rectify issues, ensuring robust code outputs with minimal debugging.
Addressing Common AI Challenges through Precision Prompts
Common issues in generative AI outputs, such as generic content, inaccurate facts, or inappropriate length, are typically rooted in prompt imprecision. Clearly stating the intended audience, specifying desired length ranges (e.g., “approximately 100-150 words”), and explicitly requesting verifiable citations significantly mitigates these problems.
Additionally, because AI models have limited token memory, users should strategically structure complex prompts into smaller, sequential requests. Reminding the AI of critical context periodically in extended interactions prevents information loss or drift in model responses.
Ethical Best Practices and Responsible AI Usage
Generative AI introduces ethical considerations library workers must proactively manage. Prompts could explicitly instruct AI to avoid biased outputs (“Highlight and avoid cultural, gender, or age-related biases”) and always require users to verify AI-generated content independently. Privacy safeguards necessitate removing personal identifiers from prompts, and transparency mandates indicating AI-generated content clearly, e.g., “Content generated with AI assistance and verified by library staff.”
Cultivating Prompt Literacy among Library Staff
Promoting widespread prompt literacy within libraries requires structured and ongoing staff engagement. Regular training sessions and hands-on experimentation, such as brief, informal workshops, can encourage staff to collaboratively refine challenging prompts, evaluate outcomes, and collectively learn from both successes and setbacks. This approach fosters continuous improvement and builds practical expertise across library teams.
Conclusion: Towards a Prompt-Literate Library Future
By adopting precise, structured prompt-engineering practices, library workers can transform generative AI from novelty tools into powerful collaborators. CLEAR and PTCF frameworks, iterative prompting strategies, and responsible ethical practices ensure that AI tools reliably enhance library services.
Investing time to develop robust prompt literacy among generative AI users enables consistent production of accurate, relevant, and ethically sound outputs, freeing library workers to concentrate more on the strategic, innovative, and distinctly human aspects of their professional roles.
In Practice
Designing AI Tools for Higher Education
Libraries are partners in creating generative AI tools for higher education. Clarivate’s product development team discusses why community engagement is at the heart of what they do.
Click here to read more about how the library-vendor partnerships are building a better generative AI future.
Library technology and content providers understand that to build trust around AI, librarians must have a seat at the table. Many thanks to Christine Stohn, Marta Enciso, and Cristina Blanca-Sancho from Clarivate’s Academic AI team for discussing how Clarivate includes key stakeholders in generative AI project development.
How does Clarivate invite librarians and library workers to participate in the development of generative AI tools?
Community engagement is really at the heart of how we develop AI tools at Clarivate. We actively involve librarians and library professionals through advisory councils, beta programs, and development partnerships. A great example is our Academia AI Advisory Council, which includes librarians and academic leaders who help guide the direction of our AI use and tools—offering input on use cases, priorities, and ethical considerations.
We also run regular webinars—a recent one talked about AI output evaluation—and sessions at conferences that feature library-focused discussions, such as the ones taking place during Clarivate’s Directors’ Days, the Charleston Conference in the US, the UKSG in the UK, or at the ELUNA Ex Libris user group conference. We also publish surveys like the “Pulse of the Library” or direct consultancy calls with librarians for specific projects to gather insights directly from the field. These feedback loops ensure we’re not just building with AI, but building the right tools that truly support the evolving needs of libraries and their communities.
Clarivate is using a suite of AI models for research, analytics, and metadata creation. Do users need to use different prompt techniques to interact with each of these AI applications? Or are there general guiding principles for prompt engineering?
While each of our AI applications has its own strengths, we encourage users to follow a few core prompt principles across the board—things like being clear, specific, and providing enough context. These fundamentals go a long way, regardless of the tool.
That said, one of the great advantages of large language models that Clarivate fully leverages is that they support natural language input, which means interacting with them is becoming more intuitive. You don’t need to master complex syntax—just describe what you need in your own words.
With the rise of Agentic AI, we’re seeing a shift in how users interact with these systems. Instead of issuing step-by-step commands, users are encouraged to describe their goals—even complex ones—as they would to a colleague. For example, rather than building a detailed Boolean search query, you might simply say what you are trying to achieve, and the agent will handle the rest—planning the steps, selecting the right tools, and executing the task.
So really, the goal isn’t to make users learn more complicated prompt techniques—it’s the opposite. We’re lowering the barrier so users can focus on what they want to achieve, and let Clarivate’s tools and the Academic AI Platform behind them take care of the complexity behind the scenes.
What do engineers think about when they are creating tools specifically for higher education? Are there standards for UX design for generative AI applications?
Having a dedicated UX team is absolutely essential when building tools for higher education. It’s a specialized area, and our UX experts play a central role in making sure the tools we create are not only functional but also intuitive and aligned with the needs of students, faculty, and librarians. They regularly conduct user research—surveys, interviews, usability testing—both before and after launch, to ensure we’re designing with real users in mind.
At Clarivate, we follow clear design principles: transparency, academic integrity, accessibility, and ease of use are always top priorities. When it comes to generative AI, UX design also emphasizes explainability and user empowerment. We want users to clearly see where AI is being used, understand how it works, and feel confident integrating it into their workflows.
Of course, designing for higher ed isn’t always simple—there are a lot of different needs and contexts to consider—but these are the standards we aim for. Ultimately, it’s about empowering users without overwhelming them.
You recently hosted a webinar on evaluating the quality of AI output. What advice can you give librarians who are evaluating the use of AI tools?
In the webinar, we focused on evaluating AI output using clear criteria like accuracy, faithfulness to source material, and overall usefulness. At Clarivate, we use a detailed set of internal benchmarks, and I always recommend that librarians take a similarly structured approach—don’t just ask “Do I like this?” or “Is it useful?” Instead, work from a defined list of criteria.
Beyond the basics, and when answers are grounded in multiple documents from our indexes, consider things like context relevance—are the documents used in a retrieval-augmented setup actually relevant to the query? Do they reflect diverse viewpoints? Is the output comprehensive and appropriate for the academic context? It’s also important to test how the tool handles low-information queries, where a strong answer might not even be possible. If looking at individual documents, does the output correctly represent the views of the author? Are the facts extracted from the document accurate?
Ultimately, every library will have different priorities, so it’s about identifying what matters most to your users. Manual testing can be time-consuming, so focusing on those priorities is key to making the evaluation process manageable and meaningful.
AI hallucinations and bias are known issues. How do you minimize the risk of hallucinations, bias, and other data errors?
Minimizing hallucinations, bias, and data errors is a top priority for us. One of the key ways we address this is through retrieval-augmented generation, or RAG. This approach grounds AI responses in curated, trusted content—so instead of relying solely on what the model learned during training, it pulls from reliable sources like ProQuest, Web of Science, and the Ex Libris Central Discovery Index.
Diversity in our data sources has been a long-standing focus at Clarivate. It’s not just about volume—it’s about ensuring a wide range of perspectives are represented, which helps reduce bias and support more inclusive outcomes.
On top of that, we design our system prompts with strict guardrails to prevent unsupported or inappropriate outputs. And we test extensively internally and listen to our customers’ feedback. The testing includes manual testing and increasingly, automated checks. It’s a continuous process, but it’s essential to ensure the AI behaves reliably and responsibly.
What is Agentic AI and how can it help scholars and library workers in doing their jobs?
Agentic AI refers to intelligent agents capable of reasoning, planning, and autonomously executing complex academic tasks. Unlike traditional chatbots, these agents can carry out multi-step workflows. We’re already applying this technology in tools like the Web of Science Research Assistant for literature reviews, and soon on ProQuest, where users will be able to chat directly with documents. In our research intelligence solutions, agentic AI helps institutions and researchers to identify funding opportunities, suggest collaborators, and compare institutional research output. And in library operations, agents can assist with tasks like generating and enriching metadata, managing budgets, and streamlining acquisitions. While today’s agents are still relatively simple, we see huge potential ahead. Future systems will be more advanced, capable of coordinating across multiple tools and even collaborating with other agents to handle a broader range of time-consuming tasks—freeing up scholars and library staff to focus on higher-value work.
Do prompts/user behavior need to change to work with these new, more complex algorithms?
Yes and no. Agentic AI makes it easier to use natural language—often in any language—to get meaningful results, reducing the need for complex search syntax like Boolean queries. The main shift is in mindset: instead of issuing simple commands, users are encouraged to frame broader goals or tasks. For example, rather than asking for a list of articles or entering keywords, a user might request, “Prepare a literature review on climate policy impacts in Europe based on the last 10 years’ worth of research.” That said, users must remain critical of AI output—generative systems can still produce errors, so human oversight is essential before treating results as authoritative.
Let’s talk a little bit about student success. Why is it important for students to learn how to use AI in their higher education experience?
AI literacy is quickly becoming a core skill in higher education. Students who learn to use AI responsibly can boost their research, productivity, and critical thinking. Many are already experimenting with tools like ChatGPT, so formal guidance helps ensure ethical use—like proper citation and verifying outputs before relying on them.
How can information literacy sessions set students up for success in their working lives?
Information literacy sessions—especially those that include AI—really set students up for the real world. They teach students how to organize their thoughts so that they can clearly state their goals in AI-based tools to achieve the best possible output (i.e., how to formulate their prompt), to think critically about AI-generated content, spot bias, and use these tools for personal development, research and analysis. These are skills they’ll carry with them into any career.
Reading Room
A selection of multimedia content to fit your time commitment.
5 minutes
- Generative Artificial Intelligence for Teaching, Research and Learning, UC Davis LibGuide, 2025.
- Getting started with prompts for image-based Generative AI tools, Harvard University Information and Technology, 2023.
- Must Known 4 Essential AI Prompts Strategies for Developers, by Reynald, Medium, 2025.
10 minutes
- AI Pro Tips Series: Prompt Like a Pro, by Emily Laird. Generative AI 101 with Emily Laird Podcast, 2025.
- Prompt Engineering, from The Center for Digital Learning and Innovation (CDLI) at Ohio State University, 2025.
- What is prompt engineering? Video by Lindsey Mercer, University of South Florida, 2024.
20+ minutes
- AI literacy and its implications for prompt engineering strategies, by Nils Knoth et al., Computers and Education: Artificial Intelligence, June 2024.
- CLEARer Dialogues with AI: Unpacking Prompt Engineering for Librarians, webinar with Leo Lo. Choice-ACRL webinar, 2023. Video below.
- Evaluating the quality of generative AI output: Methods, metrics and best practices, webinar with Christine Stohn and Marta Enciso, Clarivate, 2025.
- Prompt engineering in higher education: a systematic review to help inform curricula, by David Lee and Edward Palmer, International Journal of Educational Technology in Higher Education, 7 (2025).
Quiz Time
Test your AI literacy knowledge!
Discussion Questions
AI gives us a lot to think about. Share your thoughts in the comments!
- What are your prompt engineering trips and tricks?
Coming up
Next week, Juan Denzer, Engineering and Computer Science Librarian at Syracuse University, writes about evaluating generative AI resources. And we investigate how SUNY Empire State College is making collection development decisions around AI.
Take me to the Table of Contents.
Know someone who might be interested in this micro-course? Share the registration page with your colleagues! Make sure you’re registered to receive all eight weeks directly in your inbox.
Learn more about Clarivate and Choice’s LibTech Insights.
119 responses to “Week 4: Prompt Engineering”
-
Read and understand first before using AI.
-
What are your prompt engineering trips and tricks?
I’ve actually been working on a course workshop (or two) for the fyc course I teach at my other university where I am a graduate teaching instructor. So far, it follows closely to the two frameworks mentioned in this week’s reading where I ask students to think of the genAI’s persona that they want, specificities about what they want it do for them/the assignment, and being specific about the kind of output that they want. I am still working at it for this coming fall semester, but that’s what I have so far as “tips and tricks.” -
What are your prompt engineering trips and trick
I think it is better to read and understand first before you start with AI. Be specific about what you want, and
provide relevant background or source material. -
When using AI or prompt engineering, the statement or question should be straight to the point to receive real-time responses.
-
The information about roles/personas in prompts is particularly useful!
-
What are your prompt engineering trips and tricks?
So my prompt engineer trips that I have is starting off who you are a person for example I would tell the AI you are a librarian…. Or you are a medical student….. this seems to help to give the AI response back to fit more of what is needed.
-
Because I work in the library, my prompt engineering trick would be to ask AI as a middle school librarian, then give AI a library scenario.
-
I try to be precise and provide complete sentences. I try to get the details by asking AI in detail, informing the format I require.
-
Having many years of experience with the reference interview has been helpful when crafting prompts, because I think of it kind of like unpacking a research topic in order to approach clarity. I don’t always have a firm outcome in mind of where I want the AI to take me, but I do stay aware and vigilant so that I can spot bias and hallucinations when needed.
-
The discussion here makes prompt engineering feel approachable rather than technical. It highlights how thoughtful questions can guide better learning outcomes.
-
I usually will ask short and clear questions and include the format and length of the response. Through trial and error I learned that vague questions led to responses where I spend a lot of time refining the responses to get better answers. I also noticed the tool drifting from the original question when not given more explicit directions. So now I also include a persona and tone to my prompts.
-
I will ask the AI to be in the situation for example : Imagine you are the customer service librarian….. I will give as much as detailed prompt in numbering, step by step of instruction.
-
I follow a step-by-step approach: after each step I evaluate the response and then ask a targeted follow-up to refine it. My prompts are precise and include the necessary context to guide the outcome. I aim to be clear, exact, and thoughtful when defining what I want.
-
When writing a prompt I always come up with some full scenario in mind. That way I can shorten it and give it to the AI for background information. Such as, I’m having to write a email giving bad news to someone; I come up with a scenario that maybe something like Jake’s cat died due to…etc. etc…and his mom is emailing him this news. I would then just give the AI a prompt that is just Jake’s cat died and mom is emailing him. I will then use inspiration from the wording that the AI gives to write my own email about bad news.
-
I provide as many details as I can. Usually, I use it only to refine work that I have already done.
-
Be clear in the prompt, move forward as I get answers, always based on the answer and taking advantage of the suggestions I receive. I try to write in English, it’s more effective, but I’m not fluent, so I also write in Portuguese. Asking for confirmation from all sources.
-
I try to be as specific as possible when asking questions and refine the prompts at every stage. Asking clarifying questions, limiting the work count and specify what format I want the information presented is helpful. I tend to assign the AI tools a role. By experimenting with the different tools, I have found the quality of the responses improves when giving the tool a persona. The answers become more focused and not so vague.
-
I use a Prompt Engineering Framework to craft ‘precise’/specific commands. Adjust as necessary.
-
When it comes to prompt engineering, I’ve found a few things that really help. Being specific is important. You should say what you want, the context, and the format. If it’s a big task, I found that breaking it into steps works better than asking all at once. I also like to ask the AI to act in a certain role. And honestly, it’s all about trial and error. You often need to tweak your prompt based on the response you get.
-
I still have alot to learn about prompting properly.
-
De lo que aprendido y practicado es incluir palabras como dar una orden completa de lo que necesito, utilizando prompts claros y con estructura
-
My AI trick in getting the best of what I need is to be specific and provide enough information while still allowing for creativity. On the other hand, I provide my expectations for the output. Lately, I have discovered that keywords are very powerful and when keenly attributed, they provide a valuable input depending on the topic prompted.
-
Quick documentation → Summarize technical standards or complex papers to speed up understanding.
-
I use a step-by-step process, by evaluating each response given at the end of every step and then asking a refining question. I keep my prompts precise . I also provide necessary context to steer the prompt.
I try as much as possible to be clear, precise and profound with definitions of what I want. -
I try to be clear, specific and to provide context
-
Cuando necesito, suelo solicitar varias opciones para elegir la mejor, cuando es un trabajo extenso ordeno las tareas paso a paso y aumento más texto, hasta lograr la respuesta deseada.
-
I provide context and try to give as much information I could. Then I review what AI gives me and add/take away what I need to to generate the most relevant response.
-
I would say to be precise, clear and also provide context
-
I like to start broad and then iterate step by step. First I ask for an outline, then I request detailed content. This way, I can adjust before the AI goes too far in the wrong direction.
-
To get the best results from an AI, be specific, provide context, and use clear instructions.
-
On ChatGPT I like using a new chat whenever I am asking a question on a new topic so it could answer my question fresh without any previous references. Sometimes I would ask the same question twice in new chats to see if I get differing responses. The strategy is helpful for me in addition to creating a more detailed prompt.
-
I’m very satisfied with this education as it gives me the whole picture of using AI as a librarian. Now I know more about AI prompt engineering.
-
I don’t necessarily have any helpful tips or tricks but I did find the CLEAR format helpful. I have had answers come back that seemed a little questionable and I can see, after reflecting on the CLEAR model how I could improve the output by removing some of the more human related language elements. My biggest struggle is that LLMs and AI are changing so quickly that it is hard to keep up. My Prompt Engineering classes that I took last fall are pretty much obsolete at this point.
-
I haven’t been using AI too long, so I am learning as I go about crafting precise but detailed prompts. It helps when AI gives me additional prompts after my first entry to refine my results.
-
I use natural language that excludes any unnecessary language to generate a clear prompt. I also will indicate if AI did not give me a correct response and restate the prompt, so evaluating responses is important.
-
What are your prompt engineering trips and tricks?
I usually make my prompts as specific as possible, while also keeping them concise. I find that adding too many clauses in an attempt to by hyper-specific results in messy output. I also like to take time to really think through my prompt so as to minimize my use of any generative AI tool. -
What are your prompt engineering trips and tricks?
Use the CLEAR and PTCF method to obtain the best results. -
I try to be as concise as possible, then ask follow up questions as needed.
-
AI Prompt-writing is actually a component of the technology I enjoy, especially as it incorporates prompt chains to refine answers. I find that the prompting component of AI usage is also a way to maintain student’s critical thinking and reading skills as they are presented with learning activities that build skill in these areas to create effective prompts. Another strength of the prompting feature is that if done well, it requires a fare amount of reasoning ability to create a prompt that would render a relevant and accurate outcome. Human review is a requirement in this process. This is another way that I can use AI prompt-writing to educate students – showing them that extracting effective information from AI is a process rather than a single or few sentence request.
-
What are your prompt engineering trips and tricks?
Clear, logic, explaine, adaptative and reflexive. Specific, and clear objetive. -
Utilizó la metodología smart para iniciar a diseñar mejores prompts
-
I think carefully crafted prompts specifying all the components (persona, goal, task, and output format) is becoming less necessary, but iteration is still important to refine the results.
-
Be as clear and specific as possible. Provide context.
-
Keep it simple and concise. Check sources but be sure to research and learn on your own too.
-
I use Leo S. Lo’s CLEAR framework to standardize the method for creating optimal prompts:
The five components of this framework are:
Concise: Instructions should be delivered in clear, specific language, avoiding unnecessary information.
Logical: Message prompts should be structured and coherent.
Explicit: To receive a stronger response, instructions about content, format, or audience reach must be precise.
Adaptive: We must create different prompts about what we are presenting so we can obtain different responses from the AI.
Reflective: Continuously evaluate and improve prompts. -
Write short and clear commands. Refine where necessary.
-
Write short commands. Always request links to the source material for the answers generated.
-
Provide as much context as possible and ask the tool to ask clarifying questions before responding.
-
What could be my important prompt engineering trips and tricks?
First, establishing a structured model to ensure AI-generated content meets academic standards for reliability and relevance.
Second, as we read in several papers presented here, being specific and clear about my desired outcome.
And finally, breaking complex requests into smaller steps. -
In general: ask to include citations, be clear when making a question/asking something, ask for different scenarios, check all references!
-
Be as clear and concise as possible. Always consider the AI’s limitations. Explore different prompting techniques
-
I was fortunate to attend one of Dr. Lo’s early sessions on prompting AI and have shared the CLEAR strategy with other librarians, faculty, and students. It helps my students move from the generic types of questions they would use with Google (which, depending on whether they’re logged in or not, may have a significant profile of each user based on their interaction history) to a more precise search that would generate output at the level they are looking for.
-
Building AI literacy is very important for our institution.
-
Normalmente los que utilizo son los siguientes:
*Ser muy CLARO al solicitar contenido/información.
*Adaptar la información generada a mis necesidades.
* Solicitar citas.
*Evaluar la información recopilada. -
I’ve seen videos about people telling LLMs “who they are” in their prompts. I didn’t know why people were doing this but now I know they’re likely abiding by PTCF prompting. Similar to CLEAR, I have been taught by colleagues to make my prompts very explicit and concise, not assuming the LLM knows *exactly* what I’m talking about to get the best output.
-
The topic is new to me and I’m learning a lot.
-
I really liked the part of motivating librarians to learn how to use and do prompts for generative AI. It be a disservice for our patrons if we don’t learn about this in order to guide them in the most ethical ways possible.
-
Some prompt engineering tricks I use are as follows:
1) Being very CLEAR when asking for content/information.
2) Asking the AI tool to expand and explain specific points generated.
3) Tailoring the information generated to suit my needs.
4) Asking for citations.
5) Evaluating information gathered. -
I strive to be concise and descriptive.
-
What are your prompt engineering trips and tricks?
Just to echo a lot of other folks on here — be as specific as possible. Don’t take AI at face value, but write your prompts in such a way that you can brainstorm or breakdown a topic versus looking for concrete answers. The more you know about a topic, the more you can use AI as a resource to research that topic. Never let AI be your only resource.
-
I have learned to get very specific from the start so I don’t have to do multiple attempts. I have to keep in mind that AI is actually not smart at all and needs very good instruction. It doesn’t actually “know” anything, even though sometimes it seems like it does. It can’t figure out things a human could. For example, I asked it one time for an image of a dragon breathing fire; no matter what I put in for a prompt the fire was always coming out of the dragon’s chest. A human would know that isn’t what I want. I think we need to remember that AI is not logical.
-
Some of my prompt engineering trips and tricks include but not limited to: been as Specific and Concrete as possible, this involves clearly stating what you want, how you want it and in what format; Giving the AI tool clear role and context; Using step by step instructions; Using chain of thought prompting and Iterative Refinement, among other prompt techniques.
-
I have been asking my self how to get the best out of AI. This question has been answered after learning more about Prompt Engineering. As Librarians were too used to the old way of searching, using the Boolean tools, now this can be replaced with Prompt engineering to get answers to more complex searches.
-
For effective prompt engineering, be clear and specific about your request, provide examples or templates if possible, and break complex tasks into steps. Assign the AI a role to güide its perspective, specify tone or style, and don’t hesitate to iterate, refining prompts usually produces better results. Asking for multiple options can also give you more useful outputs.
-
Before starting this micro-course, I’ve used elaborate key phrases to be as specific as possible. For example, I recently asked copilot for “papers showing the value of pairing information literacy and scholarly communication in academic libraries.” While I got four useful sources, including an ACRL while paper, a book, a research paper and a scholarly article, I can’t help but wonder what the results would have been if I had framed my prompt using either one or both of the suggested frameworks (CLEAR & PTCF).
-
The section on ethical best practices is so timely and necessary. It’s refreshing to see an article that doesn’t just focus on the ‘cool’ aspects of AI but also addresses privacy, bias, and transparency. These are critical considerations, especially in the library world, where trust and accuracy are paramount.
-
I don’t really have a set of tips and tricks, but I try to be as specific as possible. Often I need to refine what I’m asking for. In my case, I think I learn by necessity.
-
Reading through this module, I kept thinking about how elementary teachers explain instruction and process to their students. A very popular example is asking students “how to make a peanut butter & jelly sandwich.” If students are explicitly saying exactly how to make the sandwich, you could end up with the peanut butter on the side of the crust or on top of the bag of bread. You really need to be as detailed as possible, almost like you are talking to a child.
-
To be honest I’m not used to apply prompt engineering on my AI queries. Today I tried with a role playing strategy. I asked AI to pretend to be a reference librarian who writes an email asking a profesor to evaluate our service. It helped a lot, the email content was professional, polite and concise.
-
I focus on being transparent, specific, and structured, while providing examples and breaking tasks into clear steps.
-
As an academic librarian, my main prompt engineering “trick” is to always apply a structured framework—most often PTCF (Persona, Task, Context, Format). Defining the AI’s role (“I am an instruction librarian…”), clearly stating the task, adding any necessary background, and specifying the format prevents vague or unusable outputs.
I also use iterative prompting: starting broad, checking accuracy, then refining for depth or clarity. This works particularly well for complex topics, where breaking the prompt into smaller, sequential requests ensures the AI stays on track and overcomes token length limitations.
For citation-heavy outputs, I explicitly require verifiable references with DOIs and always cross-check them in authoritative databases to avoid “hallucinations.”
When generating images or multimedia, I include detailed descriptors—style, mood, perspective—to align results with our library’s instructional goals.
Finally, I embed ethical safeguards in prompts, such as instructing the AI to avoid bias and excluding personal identifiers, while making clear that content will be verified by library staff.
Prompt literacy is a skill we continue to develop collectively in the library, learning from both successful and failed prompts to improve efficiency and reliability.
-
Very interesting to apply prompt for exact information.
-
What are your prompt engineering trips and tricks?
I am not familiar with prompt engineering, but based on my limited experience with AI, I believe that providing concise facts is necessary to generate helpful results.
-
I don’t have much experience with using AI tools, however, what I have learned in my limited experience is to be explicit with what I want/need. So if I’m putting together information for introductory level students then I need to include that info in the prompt. Or if I’m coming up with an explanation sheet for faculty then that needs to be indicated to the AI tool. What the info is for is very important, since the audience of the product matters as well as the content.
-
I’m still developing my prompt engineering knowledge. When I do use AI I try to be as concise as possible to mitigate the number of times I have to hit the enter button on the platform. I think it is essential to have everyone learn prompt engineering to help make AI a more effective tool.
-
Reading through all the comments has been educational. I agree that prompt engineering will probably become less important in the future as AI models develop and improve. However, knowing the basics about prompting is as important as having basic information literacy skills. There is always something to be learned, tweaked or improved on.
-
I try to picture exactly it is I want to design, so I can use the correct wording.
-
To be honest, I haven’t engaged with generative AI enough to really identify the best prompt process. It’s something I’m still actively learning. What I have learned is that you need to be specific and you need to correct it as it answers…and you need to know about what you’re asking about because it lacks nuance-you will need to bring your own. On the other hand, don’t try to “trick” it, because it is very very very good at guessing, which is its function.
-
Exactly, your last line.
-
-
What are your prompt engineering trips and tricks?
For me – now that I have gained knowledge on prompt engineering the first trick I bear in mind is the precision followed by tokens for AI models. The latter are essential because you instruct AI model to give you the exact information you are looking for, followed the fact that the tokens of AI models are very limited.
The CLEAR framework (Concise, Logical, Explicit, Adaptive, and Reflective) has also helped to instruct AI model (ChatGPT in particular). This framework eliminated vague prompts/outputs. Furthermore, the framework saves time between me and AI model I and using and also cuts the unnecessary prompts be it I’m asking for image creation or text-based information.
For examples rather that prompting AI model to search “AI era and information literacy” I prompted it to “Kindly search for me 10 peer-reviewed articles on the role of information literacy in the AI era. These articles should be published from 2023until now”. The model took my instructions and gave the exact articles (20230 until 12 August 2025 with links to full-text of each article and short summary. I further instructed the model to give me the Open Access PDF’s from the list it has extracted, the model did just like that. I was able to save time, get exactly what I have asked for within a very short space of time. As a librarian this is critical for our researcher unlike when the search the actual online databases for their research. -
I am always clear and provide as much context as possible
-
Well, I like concise and logical; minimalist.
-
Some of the prompt engineering sounds a lot like how I conduct research. When I am searching a database, I start broad to see what is available and I narrow down, getting more focused.
For students, most academic research has an audience that does not need to be named– the INSTRUCTOR. When doing prompts you need to be aware and mention the audience and style of the text being generated. Prompt engineering makes you more aware of those “hidden” academic factors. -
I try to provide as much context as I can and be specific about the format of the output suggested. I also tend to ask follow-up questions to modify the results. I don’t always think of everything I need initially in the prompt to get the best results, so appreciate the chance to follow up.
-
Make AI search the web for citations so it is less likely to rely on its internal knowledge
-
What are your prompt engineering trips and tricks?
It is an AI tool that can encourage staff to collaboratively refine challenging prompts, evaluate outcomes, and collectively learn from both successes and setbacks. It requires evaluation of the tools to ascertain there accuracy ,faithful source and usefulness . The trips and tricks of prompt engineering trips include replacing simple commands with complex ones and paraphrasing the results before using them . -
depend on what i’m looking for. CLEAR is always the goal. using designed prepared prompt is the same as limiting to a field in traditional databases.
-
What are your prompt engineering tips and tricks?
Drawing from Lo’s CLEAR framework, starting a prompt with Concise commands is critical. It’s important to remove any unnecessary words to ensure preciseness in the prompt. A helpful tip might be to first formulate a prompt query or command in writing before entering it into the GenAI tool. Thinking about a good and concise prompt with precision will promise better accuracy in the AI’s output.
-
When learning to write strong prompts it helps to think of it as a conversation with a knowledgeable assistant who knows nothing until you explain it. Be clear so the response will be useful. One way to do this is by giving the model a role such as historian marketing expert or teacher. Setting a role helps shape voice style and focus.
Begin with a clear action using a strong verb like summarize write create compare or explain. Provide just enough background to make your request clear without unnecessary detail. Imagine briefing a colleague and include only what they need to start.
Tell the model how you want the answer presented. If you want bullet points a short paragraph or a table say so. Set a length limit to keep the answer focused.
Prompt writing often works best in small steps. Start with an outline or list of ideas then follow up to expand the parts you care about. Once you have the content ask the model to adjust the style or shorten it.
Use examples when you need a specific structure. One or two sample inputs and outputs can help guide the model. If the first attempt is not right refine your request. Small changes in wording or detail can make a big difference.
You can also control creativity. For single correct answers keep the creativity setting low. For brainstorming or idea generation increase it. Keep experimenting and treat prompt writing as a skill that improves with practice.
-
Iterate and refine prompts for better results. A request for authoritative sources is needed.
-
I try to follow the CLEAR guidelines, but I also like to see what it does with a super vague prompt. I tend to start simple so I can experiment with prompts to see how it changes.
However, some of the recent discussion about the upcoming release of GPT-5 makes it seem like spending a lot of time teaching people how to write prompts might be wasted time. As models improve, they are generating better outputs with “bad” prompts. -
This might not sound smart, but one thing I noticed is that it pays off to present the same prompt to various models and then use you ‘human’ intelligence to decide which model “understood” me better.
-
When I design a prompt, I try to avoid ambiguous words, or I use explicit phrases focused on actions, names, places, time and roles. However, sometimes I need to read the first or second AI’s answer to improve my prompt.
I think it is like the old searching strategies, but now, we need to be more explicit. -
What are your prompt engineering trips and tricks?
I have found that I begin my prompts with an action verb, direct object, audience, and situation/context. For example, “Generate a list of 15 research questions for a 7th grade ELA class based on reading I Am Malala by Malala Yousafzai.” It seems to be more effective if I work in qualifiers, like “15 questions,” and I usually ask for more than I need so that I can choose the ones I actually want to use. After that opener, I will add in other items that I need as additional outcomes, and I might say, “Incorporate the previous directions, and include/expand on/provide… It doesn’t seem very flashy, but it saved me some mental effort at times.Learning more about the growing energy concerns has me questioning whether or not that was too high a cost for the bit of ease.
-
Using an AI I need to think on what would best help me understand what I’m having difficultly understanding. Yes AIs are not good at math but say I’m writing a paper and I need feed back on the ideas I have I first would upload the requirements my professor is asking for then add my own ideas and ask to see if what I have makes since. If it does then I can continue going back and forth understanding what could be refined or looked over again because a phrase or information is not making since. I want to think the AI as a learning partner or aid not a oh this will solve my whole problem. So if a student says hey I need to bounce some ideas off of someone but everyone has their own thing going on do you have any suggestions. Then I can say hey this AI can help bounce ideas but remember this won’t answer your problems but it can guide you .
-
Asking the AI system to take on specific personas relevant to the task at hand has really been beneficial, as well as telling it when appropriate to ignore certain sources or keywords so the specific generative response is actually relevant. I like the idea of doing a Q&A style conversation with a system, but as others have noted, the environmental impact of the data centers fueling these LLMs leads me to use the least amount of prompts as possible to get whatever I necessarily need from them, which I also try to strategically minimize. As a librarian, I feel I have a larger role to play as someone who is well-versed in how these systems work and related issues, not necessarily by being a power user of any of these AI systems and contributing to these ongoing environmental issues.
-
It strikes me that using AI chatbots effectively (if you feel that’s valuable…) would be well-supported by the skills librarians already teach re formulating search queries and identifying keywords. For all the gloss of novelty from the AI boosters, there’s a lot of overlap. Maybe to disguise the amount of time you have to exert trying to get the tool to produce what you want.
-
I prompt AI by using complete sentences. I also use please when asking and thank you when ending something. Kind of like talking to a person even though it is on a computer.
-
While I don’t usually like encouraging people to engage with these technologies if there are better and more efficient options out there, let alone encourage multiple queries due to the undue and increasing burden on the environment for each output. If someone is insistent on using these technologies to do research work, I do recommend that they prompt a chatbot that has access to the internet to check to confirm their first results against the internet to ensure the title really exists with that author and that the URL/DOI is accurate. Honestly, I find that the URL and DOI still can be pretty wrong but the odds of a title actually existing with the stated authors has a higher probability of being correct. Unfortunately, from what I’ve heard, the real way of actually getting better results form your prompts is to pay the rediculous sums asked by these companies to access “more advanced” models but I do not have the resources nor the desire to check the veracity of these claims (and it appears according to community discussions in Reddit that some of these more advanced models may be increasingly producing poorer results and users may not be feeling that they are getting the same amount out of their investment that they once were).
-
Start with as much specificity and clarity as possible. Generally, when you provide more detail, you will receive more accurate and relevant responses. In addition, set the role or context so the AI knows how to be in the correct tone and focus (e.g., “You are an experienced academic librarian” or “Explain it like I am a first-year university student”). When asking the AI questions or prompts, try to limit your request to a smaller number of elements or steps so it’s clear. Include sample or example formats so the AI understands what you expect. Similarly, you can include constraints, for example, a number of words, required format, or tone, which will create a more focused response. If the first response is not quite right, you can refine or rephrase your request, or ask the AI to extend, condense, or reorganize.
-
This is definitely not a new tip by any means but we tell students not to prompt too many questions/queries with differing topics or themes in the same chat, as a way to help promote clarity. We also have students try out several different Gen AI tools so they are familiar with the how each differ slightly in output so they get a feel for which tool has the best potential per task. We also caution students to always refine and critique the output they receive as well by having them ask “What were the strengths and weaknesses of the Gen AI tool?” We also always let students know that AI still has potential to hallucinate and fabricate information so they always need to evaluate the output for validity.
-
A good prompt is key to getting useful and accurate answers from AI. Clear, specific prompts help guide the AI, while vague or overloaded ones can lead to confusing or incomplete responses. It’s also important to avoid sharing any personal identifying information in prompts to protect privacy. When dealing with complex questions or token limits, breaking prompts into smaller parts works best. Ultimately, thoughtful prompting combined with critical thinking helps users make the most of AI tools safely and effectively
-
This may get better as context windows expand, but working with longer documents requires iterative prompting. I used Poe.com to turn my CV into a narrative for my promotion portfolio, but I had to do it in sections. It was nice though because I could tell it to focus on a particular bullet point if it hadn’t included it. I recently tried using Copilot to write database descriptions, and I had high expectations that had to be dashed. My first attempt involved an Excel spreadsheet, and I didn’t realize that Copilot wasn’t retaining the previously added database descriptions. Instead, it would generate a new Excel file with descriptions for just the databases it queried and generated. Maybe if I had told it to use the previously generated file each time it would have gone better, but I defaulting to using it to write the descriptions in chat. I have had worse experiences, thus far, with image generation, despite being very specific about the outputs. My last attempt I hit the wall at four prompts to get a thought bubble to detach from a character’s head instead of being behind it, etc. So, I don’t know at this point. I think the study on AI slowing professional coders down 20% has merit because if we have to fact-check it all the time, what’s the time savings? Is it worth the cost of heating up the Great Lakes? (I just found out about the proposed Ypsilanti UM-Los Alamos datacenter.)
-
Another protocol for connecting to outside sources called MCP Model Context Protocol. https://modelcontextprotocol.io/overview It has a similar function to RAG Retrieveal Augmented Generation in that it helps reduce hallucinations and makes the content more accurate in AI.
-
While we speak of AI ethics, what needs to be given focus is the fact as regards how scientific libraries as evolving knowledge infrastructures can play a cardinal role in structuring AI ethics.
-
Prompt engineering involves crafting clear, specific inputs to guide AI toward accurate and relevant outputs. Best practices include breaking down complex tasks, using context, and refining prompts based on results.
-
Use correct spelling and grammar. Write complete sentences.
Be clear, specific and detailed about your request to the AI.
Provide context and perspective to focus the AI output.
Break down complex tasks into multiple short prompts.
Persona identification of AI.
Specify the desired format, tone and style of the output.
Reiterate concepts for emphasis and clarity. -
To prompt engineering AI, I would like to suggest to keep in mind in adopting AI that it is better to compare between the two sources of information- the AI generated information and the other sources of Information in order to assess the validity, accuracy and quality of information.AI tools must be capable of avoiding unnecessary burden of information to generate research report.
-
I use a step-by-step process, evaluating each response and then asking a refining question. I try to keep my prompts concise (no unnecessary words) as if writing to a word or character count. I also provide necessary context to steer the prompt.
-
Although not covered here, putting in prompts uses hybrid search. This combines a weighted mix of natural language processing and boolean logic.
https://medium.com/google-cloud/hybrid-search-combining-semantic-and-keyword-approaches-for-enhanced-information-retrieval-6a7c046c89ea -
So interesting reading all of the comments. I guess I have not been experimenting or using AI enough. I have only done some basic prompting and received good repsonses. I have not done anything that complicated, but now I have some ideas.
-
I think this is one of those areas where it’s getting simpler–Ethan Mollick and his team published a paper recently about how a lot of prompt engineering tips and tricks aren’t as vital as they once were.
In the last year or so, I’ve found giving the AI a “persona” less useful, and instead giving it straightforward context about my needs and approach has been more useful. So instead of “You’re a professor who…” I say something like “My patron is a graduate student starting research in X and is looking for…” Sometimes with personas I get a little too much roleplaying and funny voices…
I’ve done some fun info literacy vibe coding–Zotero plugins, tools for working with assessment data, etc–and I’ve found it a very good approach to start with the kind of input I have (“I have a CSV file with columns for title, author, publisher…”) and the kind of output I want (“I would like a Python program to look up this URL and return the value associated with _____”), and Claude at least is usually good enough to work out the remainder with me identifying bugs. It’s not usually necessary to specify the Python version and libraries; it’s found libraries I’ve never heard of that worked brilliantly. I definitely agree with Nishan Stepak that being able to code, even a little, is a huge help for vibe coding.
-
I’m a big fan of the CRAFT method, myself. This stands for: Context, Role, Action, Format, Tone. I do tend to downplay the ‘Role’ part since my institution seems to lean toward avoiding using AI as, say, a tutor and prefers students use AI to assist in compiling and synthesizing information.
-
I prompt the AI chatbot to create a prompt for me. I think there’s meaningful limitations to how useful a skill prompt engineering will be in the future, especially now that nearly every GenAI system uses system-level prompts and other prompt-massaging techniques to undermine how effective your prompt can be in doing anything but the most mundane.
By having the chatbot make the prompt for me, I offload the most onerous and life-draining part of using AI. -
It is better to not start with AI first. Read a little bit on your subject and get some words to use that are appropriate. Once you have a little bit and understand what you are looking for start your prompting. The objective is often to start with more general prompts, then work down to the specifics. A series of three steps is often good. Also, it is good to learn a little bit of the python or sql programming language so you understand how AI is built. It will make your searching better.
-
One tip I came across and tried successfully was uploading an example of the type of document or format I wanted ChatGPT to use as a guide. It resulted in a response that was more useful and effective than previously generated. Another tip I’ve discovered is to use a famous person or fictional character to indicate the style of the response. This has greatly informed the final product and has been quite entertaining.
-
In terms of refining or seeking a more precise response, I try to challenge like a toddler – “but why?” and say ‘I don’t think that is a true bibliographic reference can you recheck’ OR ‘but I read something that seems to be the opposite to…’ and sometimes provide the URL of a relevant authoritative source.
I have had honest responses and admissions that the AI was not correct… -
I use iterative process. Ask questions, learn from the answers, refine my questions/ prompts. Add value each time to get better results until I get my desired outcome. Works fine for me.
-
One trick I’ve heard is to ask the AI to list 10 words to describe a profession, such as a car mechanic, then use those words to describe the person of that profession in framing a question, thus describing the characteristics the AI should be imparting when answering the question. For example: “As a car aficionado with a tool chest full of ratchets and gauges, who places a focus on fuel efficiency, reliability, and comfort, make recommendations for a new family sedan that will be cost effective and kid friendly.”
Take me to the Table of Contents.
Leave a Reply