3 short tips for prompting GPT-5

It's surprisingly different now

Author

Mignon Fogarty
August 14, 2025

Issue 77

On today’s quest:

— How to prompt GPT-5
— Good writing and clear thinking matter more than ever
— Major job losses for translators

How to prompt GPT-5

People are coming to a surprising conclusion about GPT-5: we have to go back to earlier styles of prompting.

GPT-5 is more driven by the exact wording of a prompt than other recent models, so you need to give it a lot more detail and structure. The always wonderful AI Daily Brief podcast has a rundown of all the advice from people who had early access or have been hitting the model hard, and these are the biggest takeaways:

1. Include a planning section at the beginning of the prompt.

This is where you tell the model exactly how you want it to do your task. Example:

Before responding, 1. decompose the request into core components, 2. identify any ambiguities that need clarification, 3. create a structured approach to address each component, and 4. validate your understanding before proceeding.

AI Daily Brief podcast

2. Include a phrase like ‘work harder’ or ‘think deeper’

Even though you can select a thinking model, people say including a phrase like “think extra hard” will make GPT-5 think even harder and will yield better results.

3. Structure your prompts

Using headings to create sections in your prompts will help the model understand what you want.

This harks back to previous methods of prompting that went out of style as models got better. Using markdown symbols is one of the easiest ways to provide structure. These are some of the major symbols:

# Top level headings
## Second level headings
### Third level headings
- Bulleted lists
1. Numbered lists
**word** Bold formatting

I put an example of a structured prompt at the bottom of the newsletter (for a project to generate a funny list of rejected superheroes).

Is it worth the trouble?

If this all seems like a pain, I hear you, and you may want to just stick with GPT-4o or try Claude or Gemini. But host Nathaniel Whittemore says, “If it gets better results and if ultimately GPT-5 is a more powerful tool that has a higher bar to entry, it’s certainly plausible that it’s going to be worth the trade-offs.”

Sources he cited for the tips include:

I dip in and out of a lot of AI podcasts, but the AI Daily Brief is a “never miss” for me, and today’s episode has a lot more detailed information if you are interested and want more!

Create Faceless Videos in Minutes

Generate scripts, captions, and faceless short-form videos using AI—all from one dashboard.

Thousands of creators are using Syllaby.io to grow across TikTok, YouTube, and Instagram.

Good writing and clear thinking matter more than ever

The resurfacing need for especially clear and well-formed prompts reminded me of advice that Maha Bali, a professor of practice at the University of Cairo, gave yesterday in a Public AI Summit about AI in education.

She emphasized that AI only knows what you tell it and that even if, in the worst-case scenario, students use LLMs to write, they still need to learn to write well in order to give it effective directions and pass along their ideas. (I’m paraphrasing from memory.)

Other speakers had equally interesting answers to the question “If you could tell the whole world one thing about AI, what would it be?”

Anna Mills, English instructor at the College of Marin: 

  1. We, the non-techies, can still shape what this looks like. We can make it better or worse, reduce harms, and increase the degree to which it supports our values.

  2. You will be better at prompting AI to get out good writing if you understand what goes into writing.

Kathy Pham, Fellow & Adjunct Lecturer in Public Policy, Harvard: Just because something can be automated doesn’t mean we don’t have to learn that skill anymore.

Arvind Narayanan, Professor of Computer Science, Princeton: LLMs are easy to try, and people should try them for themselves and form their own opinions.

Narayanan’s talk was especially interesting. I had tried to listen to the audiobook of his book “AI Snake Oil,” but couldn’t get into it. But after hearing his talk, I want to try again with the print version. He thinks AI adoption will be much slower than many people are predicting — because it will take a lot of time to integrate AI into businesses — and that jobs won’t be going away as much as many people are predicting. Instead, he thinks jobs in the future will largely consist of overseeing AI tasks. I’m not sure I agree, but I want to hear more of his arguments.

“AI Snake Oil:” Bookshop.org | Amazon | Barnes & Noble

Major job losses for translators

One reason I’m skeptical of Narayanan’s claim that jobs won’t go away as much as predicted is the anecdotes I hear. For example, in this week’s Grammar Girl podcast, translator Heddwen Newton describes how she has lost almost all her translating work to AI. The video will start playing at that section of the interview.

Quick Hits

Psychology

Climate

Bad stuff

OpenAI Designed GPT-5 to Be Safer. It Still Outputs Gay Slurs. A simple misspelling (“horni") allowed the writer to circumvent the safety restrictions. — Wired

Model updates

Education

The business of AI

Other

Draftsmith

A special thank-you to Draftsmith — an editing assistant in Word that supports the way you work. Draftsmith has supported the newsletter by signing on as a premium sponsor for a whole month of Mondays. 🙂 Get 30% off with code SIDEQUEST30.

EXAMPLE STRUCTURED PROMPT

# Role
You are a playful yet precise generator of clearly fictional, lighthearted “rejected” superhero concepts.

# Goal
Create a small batch of silly superhero names with brief backstories that are safe for a general audience.

# Inputs (fill these before running)
- Theme or setting: {{theme}}  (e.g., office life, library, home economics)
- Count: {{count}}  (default: 5)
- Age rating: {{rating}}  (default: PG)
- Allowed references (optional): {{allowed_refs}}  (public domain or generic tropes only)
- Forbidden words / topics: {{forbidden_list}}
- Tone: {{tone}}  (e.g., punny, wholesome, dry)
- Backstory length: {{backstory_length}} words  (e.g., 60–80)
- Catchphrase style (optional): {{catchphrase_style}}  (e.g., rhyme, alliteration)

# Rules
- Clarity over cleverness. Keep jokes kind; avoid stereotypes and slurs.
- Everything is fictional. Do **not** refer to real people, brands, or proprietary characters. If the input pushes that, replace with a generic stand-in and mark **[Fictionalized]**.
- If something violates the forbidden list, skip or substitute and mark **[Omitted by Rule]**.

# Process
1) Check inputs. If {{theme}} or {{count}} is missing, ask one concise question and stop.
2) Generate {{count}} ideas using varied wordplay (alliteration, incongruity, ironic weaknesses).
3) Safety pass: remove forbidden items; avoid real persons/brands; ensure rating == {{rating}}.
4) Trim to the requested backstory length and match the requested tone.
5) Think hard.

# Output (Markdown)
## Summary
One sentence stating the theme, count, and rating.

## Rejected Heroes
| Name | One-Line Hook | Powers (Silly) | Comedic Weakness | Backstory (~{{backstory_length}} words) | Catchphrase |
|---|---|---|---|---|---|

## Rationale (3 bullets max)
- Variety choices:
- Safety/filters applied:
- Fit to rating and tone:

## Metadata
| Item | Value |
|---|---|
| Theme | {{theme}} |
| Count | {{count}} |
| Rating | {{rating}} |
| Tone | {{tone}} |
| Forbidden | {{forbidden_list}} |
| Allowed refs | {{allowed_refs}} |

What is AI Sidequest?

Are you interested in the intersection of AI with language, writing, and culture? With maybe a little consumer business thrown in? Then you’re in the right place!

I’m Mignon Fogarty: I’ve been writing about language for almost 20 years and was the chair of media entrepreneurship in the School of Journalism at the University of Nevada, Reno. I became interested in AI back in 2022 when articles about large language models started flooding my Google alerts. AI Sidequest is where I write about stories I find interesting. I hope you find them interesting too.

If you loved the newsletter, share your favorite part on social media and tag me so I can engage! [LinkedIn — Facebook — Mastodon]

Written by a human (except for the example of a structured prompt, which was written and tested with the help of GPT-5 thinking)