Practical tips for getting useful results from generative AI.
This page was created with assistance from ChatGPT and Claude, primarily for language refinement, organization, and structural improvements.
Inclusion of specific tools or models does not constitute endorsement. Evaluate each tool's suitability for your own needs and stay informed about emerging alternatives.
This resource is a work in progress. Comments welcome.
Last updated: June 2026
Conceptual principles
The following principles are based on research (notably Bowman, 2023) and practice (notably Mollick & Mollick, 2023a; 2023b), as well as the author's own experience.
Use the tools and experiment. 🧪 Try several tools and start in your own area of expertise, where you can judge the output. Build a feel for where they help and where they fall short.
Invest in your prompts. Minimum-effort prompts give mediocre results, so plan to refine and iterate. Results vary even across identical prompts, so treat prompting as a conversation and talk to the model as if it were human, knowing full well it is not. I think of the models as aliens emulating humans. 👽
Before you rely on the output:
Verify it. The text reads fluent and confident and is still often wrong. Check any fact, number, or reference yourself.
Mind what you paste. On free tiers your input may be stored or used for training. Share only what you would be fine making public.
(For more on why hallucination happens, plus more on privacy and bias, see Field Notes on Generative AI.)
Practical principles
Use the latest and strongest model you have access to. Some tools automatically switch to a smaller or weaker model, between or even within chats.
Turn on Reasoning, Extended thinking, or similar advanced mode if available.
Do not spend more than 2-3 minutes perfecting the first prompt. Start the conversation, then iterate.
You can edit an earlier prompt (mouse hover) if the conversation goes wrong or gets stuck.
Do not accept the first answer too quickly. Ask for what is missing: more specificity, fewer words, better examples, sharper actions, clearer risks. For example, give the model more context in the conversation, such as your role, function, users or customers, challenges, and desired output format.
Ask the model to be concrete, practical, and non-generic. Push it toward examples, trade-offs, assumptions, and next steps.
Start a new chat if the conversation drifts, gets confused, or becomes too abstract.
Try pasting the same prompt into two or more different models and compare which one works best for this task.
Current strong general-purpose models include ChatGPT, Claude, and Gemini. For building tools or prototypes through instructions rather than coding manually, examples include Claude Code, Lovable, and v0 by Vercel.
In Microsoft Copilot, a Work and Web toggle may be present, which changes what the model draws on. Work grounds answers in your Microsoft 365 content (files, emails, chats) and Web draws on the public internet instead. Neither is universally better, so consider trying the same prompt in both and comparing the two.
(These tools change quickly. The point is to test which model is more helpful for you in a given task, and others.)