AI-Assisted Workflows · Research

Generative AI Tools

As of this writing (2026), distilled from a 2025 presentation. AI tool capabilities change quickly — revisit this page periodically rather than treating it as permanent guidance.

Definition

Generative AI tools show up in research work in three broad ways: handling tedious writing tasks (citation formatting, summarizing, translating a result for a different audience), upscaling technical skills you don’t have yet (bootstrapping an unfamiliar API or algorithm), and changing the shape of the research process itself (custom-tuned assistants, structured prompting). This page collects the practical patterns from those first two categories that hold up independent of which specific tool you’re using.

Learning Outcome

After using these patterns, you should be able to structure a prompt so an AI tool’s response is actually useful for your specific situation, and use AI tools to bootstrap unfamiliar technical territory faster than starting from nothing.

Core Structure

Persona, Context, Task — a structure for prompts that need to produce something specific. Rather than asking a generic question, specify:

  • Persona: who should the AI act as, or who is it helping? (“A student with this coursework and these interests…”)
  • Context: what constraints or reference material apply? (“…pursuing a research topic that must satisfy these course requirements…”)
  • Task: what specific output do you need? (“…suggest three research topics, with justification and the specific algorithms each would require.”)

A vague prompt gets a vague answer; specifying who, under what constraints, and toward what output gets something you can actually use.

Bootstrapping an unfamiliar API or technique, in four steps:

  1. Ask for advice on the general approach before asking for code — “what’s a good way to do X?” surfaces the landscape before you commit to one path.
  2. Ask for a working example based on that advice.
  3. Ask about anything in the generated example you don’t understand (a specific API concept, an unfamiliar function) before using it.
  4. Experiment — modify the example yourself and see what breaks, rather than treating the first working version as final.

The “how, how, how” question, for getting value out of a new AI tool quickly. Rather than exploring a tool’s features abstractly, ask it directly: “I do <your actual job or task>. How can you help me do it better?” This surfaces concrete, situated suggestions faster than browsing generic capability lists.

Worked Example

Bootstrapping unfamiliar work with a GitHub API: step 1 asks “what’s a good way to extract metadata from a GitHub repo?” rather than jumping straight to code. Step 2 asks for a script that extracts commit, pull-request, and release counts for a specific repository. Step 3 asks how to generate the personal access token the script needs, since that’s an unfamiliar prerequisite. Step 4 modifies the script to also list contributors and their commit counts — testing understanding by extending the example rather than just running it as given. The result: a working, understood script built in a fraction of the time cold research would have taken.

Common Pitfalls

  • Asking a vague question and getting a vague, generically-applicable answer — persona, context, and task specificity matters.
  • Running generated code without asking about the parts you don’t understand, so nothing is actually learned in the process.
  • Treating a first working example as finished rather than as something to test and extend.
  • Assuming AI-assisted drafting or scripting removes the need to demonstrate real understanding — the check on this isn’t a stricter AI policy, it’s being able to hold your own in a Level of Expertise Dialogue afterward.

Checklist

  • Prompts specify persona, context, and task rather than asking generically
  • Unfamiliar territory is bootstrapped in stages (advice → example → clarification → experimentation), not accepted whole from a single response
  • Anything unfamiliar in a generated result is asked about before being used
  • A new tool’s value is tested against your actual task, not just its feature list
  • You can still explain and defend the resulting work without the AI’s help