Prompting tips and tricks.

1. Give the context up front

The single biggest upgrade. A model only knows what you put in front of it, so a bare request gets a generic answer. Drip-feeding corrections afterward (“no, it’s for students”... “no, shorter”... “no, the event is outdoors”) spends ten minutes reconstructing what you could have said in thirty seconds. Say who you are, what the situation is, who the output is for, and what constraints matter. Then ask.

The same habit works for a syllabus policy, a lab summary, or a study plan. If a competent new assistant would need the background to do the task, the model needs it too.

2. Say what the output should look like

Describe the deliverable itself: format, length, audience, tone, what to include and what to leave out. “A table with one row per vendor and columns for cost, deadline, and contact” beats “organize this” every time. If the output has a job (paste into an email, read aloud in a meeting, hand to a first-year student), say the job.

3. Show an example of good

One example of what good looks like outperforms three paragraphs describing it. Paste a past email you were proud of, a rubric comment that landed well, an abstract in the style your field expects, and say “match this.” The model reads voice, structure, and level of detail out of the example far more accurately than it reads adjectives like “professional” or “engaging.”

4. Correct it rather than starting over

When the first draft is 70% right, say what’s wrong with the 30% and keep going. The conversation is state you’ve already paid for: the model has your context, your example, and its own draft in front of it. Restarting throws that away. Be specific in the correction (“the second paragraph hedges too much, commit to the recommendation”), and it compounds.

One exception worth knowing: when a conversation has accumulated wrong turns and the model keeps drifting back to them, a fresh start with a better first prompt beats fighting the history.

5. Save what works

The prompt that finally produced the right advising-note format or the clean meeting summary is an asset. Keep a doc of your working prompts and reuse them. That habit is also the on-ramp to the next level: a saved prompt used weekly is a custom GPT waiting to be built, which is exactly what the workshop does with it.

What stopped working

A lot of circulating prompt advice was written for 2023 models and has since been tested. Researchers at Wharton ran the folk wisdom through controlled experiments, and the vendors have updated their own guidance. The results are worth knowing:

• “Think step by step” on a reasoning model. Current reasoning models work through problems internally, and OpenAI now explicitly advises against the phrase. Measured effect: marginal benefit, real time cost.
• Bribes and threats. “This is important to my career,” “I’ll tip $20.” Tested directly: no meaningful performance change. Being polite is fine manners; it tested as performance-neutral.
• Expert personas as an accuracy spell. “Act as a world-class economist” does not make the answers more correct, and a mismatched persona can make them worse. But accuracy was never the real job. A persona pre-qualifies the context: it hands the model a working set of assumptions about the audience, the tone, the level of detail, and the vocabulary before you spell any of it out, and those assumptions are most of what quality feels like. “Explain this to a first-year student” changes the answer you get in every way except its truth value. Set the persona for fit, and check the facts the same way you always would.
• Magic-phrase collections. The experimental finding underneath all of this: prompt effects are model- and question-contingent, so a habit of iterating beats any list of incantations.