The most important aspect of brainstorming with AI and fostering a Spark is to provide proper context. Tell it who you are. Upload your bio or provide a link to your website. Upload your most representative papers. Describe your goals and ambitions, both conceptually and concretely. Let it get to know you. Imagine sitting down with a mentor. You would not just say, “Give me research ideas,” or, “Here is my idea, tell me what to do.” You would provide context first. Do the same with your AI tool. You must create an appropriate sounding board before bouncing ideas off of it.

For example, a faculty member might upload representative papers and describe the broader research direction they hope to build over the next five years, then ask the LLM to suggest adjacent unanswered questions or emerging tensions in the field that are consistent with that trajectory. In this case, the LLM is not setting the research agenda. It is helping surface possibilities that the researcher can then judge and refine.

Do not let AI make decisions for you because its responses sound intelligent and well composed. It may present things that appear to be good ideas, but are they technically sound? Do they resonate with you and your goals? It is programmed to sound supportive and encouraging, so make sure it is not talking you into something you would rather not do.

At this stage, it is helpful to structure your interaction with the LLM step by step. These models often perform better when complex problems are broken into smaller, sequential parts, a process often referred to as chain-of-thought reasoning. While building Container, keep each chat session focused on one idea rather than several at once. You can provide as much context and detail as you would like, but prompts should stay targeted to that one idea or stage in the process. Think of it like writing a journal article: the Introduction, Methods, Results, Discussion, and Conclusions are connected, but each serves a distinct purpose. Your chat sessions should work the same way. If you keep the conversation focused, the LLM is more likely to return sharper and more useful responses.

For example, a researcher might have a promising project idea but not yet know whether it should become one proposal, two separate efforts, or perhaps a pilot project followed by a larger effort. They may still be deciding what belongs in scope for a three-year grant, what should be set aside for later, and whether one aim is too ambitious. The LLM can help break the idea into components, compare possible scopes, and identify what would realistically be in or out of bounds for the proposed effort. In this case, the LLM is helping organize and pressure-test the structure of the idea rather than defining it for the researcher.

Avoid bouncing between too many concepts and ideas at once or within a given chat session. AI will indeed respond, but, as mentioned above, its output will be reflective of your inputs. You may have three different thoughts in your head, but provide them one at a time and give the AI a chance to respond before moving to the next one. Most LLMs are capable of handling a sequence of instructions, but overall, if that sequence is focused on one topic or idea, the results, too, will be sharper and more informative.

Container is also the stage where an idea is especially vulnerable to drift at the project level. Even if each individual prompt is focused, repeated additions, side questions, and interesting detours can gradually expand the project beyond what is coherent or feasible. At this stage, the goal is not to keep generating more possibilities, but to decide what belongs in the project and what should be set aside for later.

AI, meaning LLMs, works best in Extension when you give it clear guidelines and boundaries. If you keep asking for more revisions, more options, or more variations, it will continue generating them. In this phase, make sure to specify what it should and should not consider, such as proposal guidelines, sponsor priorities, timelines, past meeting goals, or desired deliverables. The more clearly bounded the task, the more useful the output will be.

For example, a faculty member may have a strong core idea for a project but need to decide how best to position it for different potential sponsors. They can gather abstracts of recently funded awards from publicly available sponsor databases and ask the LLM to identify recurring features: what kinds of problems have been funded recently, whether there are common themes, what level of risk appears to be supported, and whether those patterns have shifted over the last few years. That output can then inform how the researcher frames the same work differently for different sponsors. The LLM is not deciding the strategy, but it can help synthesize patterns that support more thoughtful positioning.

Like a toddler in a grocery store, given free rein, AI will keep wandering unless you set boundaries. It is your job to keep it reined in. If you and AI are generating a meeting agenda and it is too long, tell it to make it shorter. If you ask it for a list of research payoffs, make them aligned to sponsor needs.

At the same time, in the example above, do not let patterns from prior awards dictate your intellectual direction. Those patterns can inform your strategy, but they should not push you into distorting your idea or chasing trends that do not genuinely fit your work.

One useful principle at this stage is prompt engineering, which simply means asking clearly for the output you need. A strong prompt reduces drift and ambiguity and can often get you a useful result in one query rather than several. Because prompting practices vary across systems, it is worth learning what works best in the particular LLM you are using.

By this stage, you know your work. You are the expert. You know what outputs you need, and you are capable of developing them without the use of AI. With good prompting, AI can dramatically accelerate this process. Tasks that once took weeks or months may now take a few hours. Learn about your particular AI system and the best practices for prompting, and you will fundamentally improve the Realization of your work.

For example, a faculty member may have hosted a summer outreach program for middle school students and want to describe that work effectively in a report, progress update, or future proposal. They can provide the LLM with the associated organization’s website, the summer program description, and any relevant materials used during the program, then ask it to draft a concise paragraph summarizing the program’s purpose, the faculty member’s leadership role, and the success of the effort. Because the faculty member ran the program, it is easy for them to review the draft, verify its accuracy, and refine it as needed.

At this stage, remember that you are the expert and the LLM is the tool. It can help synthesize and communicate your ideas, but it does not replace your judgment. Having a powerful tool at your disposal does not guarantee that you know how to use it well.

The model can help package and communicate the work, but it should not exaggerate the impact or invent outcomes that were not achieved. You still need to review the output carefully and make sure it reflects the work accurately and credibly.