OpenAI, the creators of ChatGPT. How do the people inside use AI every day?
I'll tell you upfront: there are no magical prompt techniques here.
"Consulting via voice during a commute."
"Taking a photo of a menu to ask questions."
It's almost disappointingly mundane.
However, they were simply the first to start these simple habits. The core principle is one:
"Consistently delegate the hassles of life and work to AI."
I will introduce 5 ways you can start copying today.
However, the fifth one is a bit different. It connects directly to the biggest change currently happening inside OpenAI.
Nice to meet you, I'm tatsuki. I support AI implementation for small and medium-sized enterprises, helping with the business integration of Claude and Codex. I am also someone who runs Claude Code all day long.
When I see people who say, "I'm using AI, but I don't feel like my work has changed much," I always think:
It's not about the features; it's about how you delegate.
You don't have to be an engineer to take these tips home. In fact, the second half is where the real action is.
▼ At the end of the article, I have prepared 7 bonuses for fully utilizing AI.
[Hack 1] Voice Brainstorming
Nick Turley, the head of ChatGPT, reportedly talks to the app in voice mode during his morning commute.
https://x.com/OpenAI/status/2075310019185389913
However, it's not to get answers. It's to organize the messy thoughts in his head by speaking them out loud.
The AI doesn't argue or rush you. That's why questions like "What's the most important thing I want to convey in today's meeting?" get settled faster than just mumbling to yourself.
I call this "wall-hitting" (brainstorming). Out of the five, this is the easiest entry point to copy.
One action you can take today: during your 10-minute commute, tell voice mode, "Let me organize what I need to do today by speaking out loud."
✗ Using it to get answers
◯ Using it to vent and externalize your own thoughts
[Hack 2] Embedding "Context" in Memory
ChatGPT has a "Memory" feature (the ability to remember context you've shared in the past). Tell it everything you want it to remember just once.
For example, just type this once at the beginning:
"I support AI implementation for SMEs. Reply in polite Japanese, starting with the main points. Please speak based on this premise from now on."
With this, all subsequent conversations will run according to your specific needs. The hassle of re-explaining your position and preferences every time disappears.
You can do the same with a "Custom GPT" (a dedicated ChatGPT where you can pre-load frequently used instructions; it takes seconds to make). For repetitive tasks, it's faster to create a "template" once.
✗ Explaining your context from scratch every time
◯ Teaching your job, position, and writing style once to make all future conversations personalized
[Hack 3] "Photo Dumping" for Indecisive Moments
Andrew Mayne, former Science Communication lead, reportedly takes photos of menus at restaurants and asks, "Which of these would suit me?"
He is completely eliminating the time spent agonizing over choices.
This isn't limited to food. Manuals, contracts, unfamiliar tables. It works best for things you've been procrastinating on because they are tedious to read.
✗ Putting things off because reading them is a drag
◯ Just taking a photo and asking, "Which part of this is relevant to me?"
[Hack 4] "Pre-cooking" for First Meetings
Mark Chen, Chief Research Officer, reportedly gives ChatGPT the background of people he is meeting for the first time to generate conversation starters and potential common ground.
He's not delegating the conversation itself. It's about raising the floor of his preparation.
For those who fear silence in first meetings, this method of "having 3 conversation seeds ready" is highly effective. It's the same for sales, interviews, or side-hustle meetings.
✗ Meeting empty-handed and enduring awkward silences
◯ Providing the other person's background and preparing with "3 common topics"
By the way, even CEO Sam Altman reportedly uses it for organizing his inbox, scanning documents, and even parenting advice. Not as a replacement for an expert, but as "another person to consult with."
That covers 1 through 4: Consultation and preparation.
The fifth one is a bit different.
[Hack 5] Delegating "Execution Itself"
While 1-4 were about consulting or asking for preparation, the fifth is about delegating the work itself entirely.
This is the usage that is currently having the biggest impact inside OpenAI.
It might be surprising, but the star of AI within the company is no longer ChatGPT.
Until a year ago, AI in the company almost exclusively meant ChatGPT. Now, about 98% of employees use an AI agent called "Codex" that can handle tasks. And it's said that 99.8% of what AI generates internally comes through Codex.
To be honest, this 98% figure is based on employee self-reporting, so take it with a grain of salt. However, the reality of 99.8% of output is likely accurate in terms of direction.
And it's not just for engineers. It has become a primary tool in departments like legal, accounting, and recruiting.
More than a quarter of the work handled by administrative staff using Codex is reportedly code-related tasks that previously had to be requested from engineers.
So, how are the insiders "delegating"? There are three principles that even non-engineers can steal.
(1) Moving from Doer to Director
They no longer do detailed manual work themselves. In the development process, humans writing code directly is zero. That is the core of the team: "0 lines of hand-written code." (Dev team details: https://openai.com/index/harness-engineering/
(2) Creating a "Place for Context" First
They don't write long instructions every single time. They prepare an AGENTS.md (an instruction sheet for agents, a single index file summarizing roles and rules) and have the AI read it every time.
The content can be as simple as a few lines:
Example of AGENTS.md content:
- Role: AI implementation support for SMEs
- Rules: Supplement technical terms in Japanese. Don't be assertive; provide primary sources.
- Tone: Polite (desu/masu). No hype.
In their words, "If you say everything is important, nothing is important." So, a 100-line "map" is sufficient rather than a thick dictionary.
(3) Small, Fast, and Frequent Iterations
The Codex team is small, with 3 to 7 people. Yet, they produce about 1,500 changes in 5 months.
The secret is "don't aim for perfection in one go; release small and fix later." It's the realization that the cost of waiting is higher than the cost of fixing.
Even if you're not an engineer, you can try this today. Give ChatGPT a goal instead of instructions.
"Do this entire task. I leave the procedure to you. Ask if you get stuck."
Rather than teaching the method in detail, give the goal and the phrase "Ask if you get stuck." This alone puts you on the "getting things done" side.
✗ Moving your hands line by line yourself
◯ Verbalizing the goal and the rules to follow, and delegating the execution itself
[Condensing these 5 into one line]
If I were to boil it down to one line, it's this:
Consult, have it write, and delegate.
1-4 are for "consultation" and "preparation," and 5 is for "execution." The more you increase the "amount of delegation" in this order, the more results will accumulate.
This isn't just a feeling. Another survey of all users ("How People Use ChatGPT," covering about 700 million people) reported that about half (49%) of ChatGPT conversations are "consultation," followed by "work" (40%), with "writing" being the most common task in a professional context. (Survey: https://openai.com/index/how-people-are-using-chatgpt/
In the end, this is all they were doing.
[Summary]
- It's not a special talent. They just started "consistently delegating hassles to AI" earlier.
- From the consultation and preparation of 1-4 to the "delegated execution" of 5. The more you delegate, the bigger the gap becomes.
- The order is always the same: Consult, have it write, and delegate.
The before-and-after is probably just this:
From being someone who ends with a single question to the AI, to being someone who delegates hassles to the AI first.
If you read this and just think "Oh, I see," tomorrow you'll end up asking a single question just like always. Most people do.
Will you try "wall-hitting" just once during your 10-minute commute today? Or will today be the same as always?
The dividing line is probably right there.
You don't have to do everything at once. Why not start with Hack 1, "Voice Brainstorming"?
For those who read this far and want to actually start moving.
I run a LINE OpenChat (you can join anonymously) where we learn AI utilization, Claude Code, and X operations together.
By joining, you can receive "7 Claude Utilization Bonuses" for free that you can use directly for work. More than 80% can also be applied to Codex.
- 300 Universal Claude Prompts
- Claude Cowork Complete Guide
- Creating Materials with Claude Design
- Automating X Posting with Claude Code
- Claude Code Starter Kit (skill + MCP)
- Claude Code Security
- Claude Code x Obsidian
How to receive:
- First, join the OpenChat: https://t.co/90omRA4UQ7
- After joining, send the keyword "Bonus" (特典).
That's how you get them.
Let's survive the AI era together with weapons in hand. If this article was helpful, please leave a review in a quote-repost and I will repost all of them!
