Advanced Skills

Overview

This page assumes you already use AI regularly and want better results than you are currently getting. It is less about knowing more features and more about being deliberate: giving the tool better material to work from, structuring requests so the output is usable without heavy editing, and building approaches you can reuse instead of starting over each time.

Work From Your Own Sources

The largest single improvement available to most people has nothing to do with prompting. It is giving the tool your actual material instead of relying on what it absorbed during training.

Paste in the document. Upload the file. Point the tool at your notes, your syllabus, your dataset, your meeting transcript. When AI works from text you supply, fabrication drops sharply, the output uses your terminology, and the result is specific to your situation instead of generic.

Some tools — Gemini Notebook among them — are designed entirely around this: they answer only from sources you provide and cite which one each claim came from. For summarizing a stack of documents, preparing from committee materials, or building study guides from course readings, that is a fundamentally more reliable approach than asking a general chatbot what it remembers.

Techniques Worth Learning

Compare “review this policy draft” with “you are reviewing this draft as a department chair who will have to explain it to faculty who are already skeptical — flag anything that will cause confusion or pushback.”

The second produces genuinely different output because it constrains what counts as a good answer. Specify who the reader is, what they already know, what they care about, and what would make the response fail.

This works for evaluation as well as generation: “read this as a student who is struggling with the material” surfaces different problems than “check this for errors.”

Asking for a finished long document in one request reliably produces something mediocre, because you have no opportunity to correct course.

Instead: ask for an outline. Fix the outline. Ask for one section. Correct the tone and level. Then ask for the rest using that section as the model.

This takes more exchanges and less total time, because you are not editing a 1,500-word draft built on a structure you did not want.

The same applies to analysis. Rather than “what should we do about this,” try: summarize the situation, list the options, give the strongest argument for and against each, then recommend. You can inspect each step, and you will often spot the flaw at step two rather than discovering it in the conclusion.

If you need output in a particular shape — a rubric, a course outline, a meeting summary in your department’s format, an assignment description — paste an example you already have and say “match this structure and level of detail.”

Two or three examples work better than one and far better than a description. Tools are much better at imitating a pattern than at following an abstract specification.

This is also the fastest way to get output in your own voice rather than generic AI prose: give it three things you have written and ask it to match how you write.

“Give me five subject lines, ranging from straightforward to attention-getting.” “Draft three versions of this explanation at different reading levels.” “Show me two structures for this presentation that emphasize different things.”

Generating options is nearly free and it prevents the common failure of accepting the first output because it is there. It also surfaces the range of what is possible, which is useful information in itself.

AI tends to be more useful evaluating work than creating it, because evaluation is a pattern-recognition task and there is an actual artifact to react to.

Useful requests:

  • What is the weakest argument here and why?
  • What would a skeptical reader push back on?
  • What am I assuming without stating it?
  • Where would a student get lost in this explanation?
  • What is missing that someone in this field would expect?

Remember the agreement bias covered in Critical Evaluation: you have to ask for criticism explicitly and specifically. “Any feedback?” produces compliments.

If you do something regularly — summarizing meeting notes into action items, generating practice questions from a reading, drafting a particular kind of message — write the instruction once, carefully, and save it somewhere you can get to it.

A reusable prompt is worth investing in. Specify the role, the audience, the format, the length, the tone, what to avoid, and include an example. It may take fifteen minutes to get right and then save you time every week.

Several tools support saving these directly — Gemini’s Gems, for example, let you store a configured assistant with its instructions and reference material attached, so you are not pasting setup text each time. Where that is available it is worth the setup.

Prompts that work well are also worth sharing with colleagues. If you build something genuinely useful, consider submitting it as a use case.

There is a point where continuing to prompt costs more than finishing the work yourself. If you are four exchanges in and the output still is not close, the task is probably one of these:

  • Something requiring information the tool does not have
  • Something requiring judgment about a specific person or situation
  • Something where you have not yet decided what you want, and no amount of prompting will resolve that for you

The third is the most common. If you cannot describe what a good answer looks like, the problem is upstream of the tool.

Before You Scale Something Up

Advanced use often means building something other people will rely on — a workflow, a saved assistant, a tool embedded in a course or a process. That crosses a line worth noticing.

Once your AI use affects other people’s work, records, or evaluation, it is no longer just a productivity choice. Review requirements may apply — around data handling, accessibility, procurement, or records retention — and they exist regardless of whether the tool was free.

Check Data Privacy and Security and Choosing the Right Tool before you build something others depend on. It is much easier to check first than to unwind it later.