GPT-6 Astra Setup Guide: 30 Steps for Business Efficiency

What should you set up when you start using GPT-6 Astra?

I researched this extensively myself.

Inference strength, permissions, Memory, browser, Codex, context...

The more I looked into it, the more unfamiliar terms I found.

I hope that by reading this article, you will be able to smoothly move on to the intermediate-level articles currently circulating on your timeline.

Now then...

Non-engineer office workers don't need to memorize everything from the start.

After looking into these 30 items, I felt the most important thing was surprisingly simple.

Decide "how you want the AI to work" before focusing on "AI performance."

This is it.

And one more thing.

The settings written on the official website are not necessarily the same as the settings actually running on your computer right now.

This time, I actually checked the contents of Codex, and I found some interesting discoveries there too.

This article is not a collection of difficult settings.

I will explain 30 items you should know before you start using GPT-6 Astra for work, from the perspective of: "If you're an office worker, why should you check this?"

*Information as of September 7, 2026. Since GPT-6 Astra has just been released, screen displays and terms of service may change in the future.

What are the "30 Initial Settings" anyway?

I'll explain just this part first.

This isn't an article about turning on 30 switches.

These 30 items are roughly divided into three types:

① Settings to check on the screen

For example, models, data usage, browser, etc.

② How to request work from the AI

Such as "how much can it decide on its own" and "what constitutes completion."

③ Settings for those who are a bit more experienced

Such as Codex's config.toml and AGENTS.md.

For beginners, the first 10 items are enough.

Let's go in order.

Level 1: 10 Items to Check First

1. First, check "where you can use Astra"

This is the first pitfall.

"I'm a Plus member, but I can't find GPT-6 Astra."

This happens. However,

"Astra hasn't come to me yet"

is not necessarily true.

ChatGPT has several entry points:

Normal Chat
ChatGPT Work
Codex

As of September 7, 2026, the entry point where you can use Astra varies depending on your plan.

First,

Check "where can I use Astra with my plan?"

This is Setting 0.

2. Update the app and Codex to the latest version

This is mundane but important.

Imagine not updating a smartphone app for months. Even if new features are added, they might not show up in the old app.

Codex is the same. In the environment I investigated this time,

Codex CLI 0.151.0-alpha.2

was running. However, a newer client was required for GPT-6 Astra support.

In the OpenAI Codex releases, a fix was included in 0.153.4 to display Astra in the bundled model picker. (GitHub

In other words,

When "Astra doesn't appear,"

before assuming "it hasn't been rolled out to me yet,"

Check if the app is old.

[Deputy Manager Taro's Measurement Note ①] Astra was missing

This time, I actually checked the model list in Codex. The result was:

11 models. 0 GPT-6 Astra.

At first, I thought, "Maybe it hasn't been rolled out yet."

But as I looked into it, the Codex CLI I was using was 0.151.0-alpha.2. This is a version older than the Astra support.

In other words,

Astra not appearing ≠ necessarily a contract plan issue.

In my case,

It was a situation where I should have suspected the client version first.

When you can't find a new AI, it's easier to check in this order:

Contract Plan

↓

App or CLI Version

↓

Rollout status to your account

3. Don't think Chat, Work, and Codex are all the same

This is also hard for beginners to understand. I roughly use them like this:

Chat

Questions, consultations, writing text.

Work

Researching, creating files, and having it finish the results.

Codex

Code, file editing, automation, etc.

For example:

"Make this email more polite" -> Chat.

"Research 10 competitors and create a comparison document" -> Work.

"Create a system to aggregate sales every morning" -> Codex.

They are all AI, but their specialized workplaces are different.

4. Being able to use Astra is different from actually using Astra

Be careful here too. Even if Astra is available, it won't necessarily be selected automatically every time.

On the model selection screen, check "which model is currently selected."

To use a human analogy, just because you hired a brilliant new employee doesn't mean that person is handling every single job.

5. Don't set "Thinking Strength" to maximum for everything

Astra has a setting to change how deeply it thinks. This is called Reasoning Effort.

Don't overthink it. Think of it like a car: "How much do I want to rev the engine?"

You don't need maximum output for correcting a short sentence. On the other hand, for work like comparing multiple documents to make a management decision, it's better to have it think thoroughly.

The important thing is that Maximum ≠ Always the correct answer.

6. Separate "work that runs on your PC" and "work that runs in the cloud"

After you ask the AI to do a job, will it keep running even if you close your computer? This depends on where it's executed.

For example, if it's a task running on the cloud side, it might continue after you close your PC. On the other hand, if it's a job operating your own computer, that environment is necessary.

Rather than memorizing difficult mechanisms, just asking "Where is this job running?" once is enough.

7. Separate Projects for each job

This is immediately useful. For example, if you talk to the AI about "Sales," "Recruitment," "SNS," and "Product Development" all in the same place, the information gets mixed up.

I recommend separating Projects by theme.

Imagine a desk: you don't pile sales materials, recruitment materials, and product planning documents all on the same desk. It's like creating drawers for each job.

8. It's better not to show the AI too many materials

I used to think, "If I give it a lot of materials, the AI will get smarter." But that's not necessarily true.

If you give it a massive amount of irrelevant materials, it becomes hard to tell what's important.

For example, if you just want it to analyze this month's sales, you don't need to give it all the meeting materials or HR documents from three years ago.

Only give it the materials necessary for that job. This is the basic rule.

9. Don't try to make the first instruction perfect

Spending dozens of minutes thinking about the instruction text for the AI is something beginners tend to do. But with agent-type AIs like Astra, you can correct the direction midway by saying "That's wrong" or "Add this condition."

So, rather than starting after writing a 100-point instruction, decide on the goal, start, and fix it along the way. This way is easier for some jobs.

10. For long jobs, check "if it will stop midway"

You asked for a one-hour research task. You closed your PC. The next morning, you saw it had stopped. This is a problem.

When entrusting long jobs, check "if it's an environment that continues even if the PC is closed" and "if you'll be notified if questions arise."

These are the first 10 items. For beginners, this is enough for now.

Level 2: 10 Items for Safety and Quality

11. If using it at work, look at Data Controls

If you use personal ChatGPT for work, this is a setting I want you to check. ChatGPT has a setting called Improve the model for everyone.

This is a setting regarding data usage for model improvement. However, there is a caution: just because you turned this OFF doesn't mean you can put any confidential company materials in. Company rules come first.

12. Don't let Memory remember everything

Memory is convenient for having it remember things like "I like this kind of writing style." But it's not a place to store customer information, passwords, absolute company rules, or confidential matters.

In particular, don't rely on Memory for "work rules you absolutely want it to follow."

13. Temporary Chat is not a "Secret Mode"

When you see the name Temporary Chat, you might want to think "nothing I write here will remain." But it's not that simple.

The important thing is that Temporary Chat ≠ You can ignore company information management rules. Before handling sensitive information, check if it's information you're allowed to input in the first place.

14. Personality is not "Intelligence"

Personality changes the AI's character and way of speaking. It's not a setting that suddenly doubles its ability. If you use it at work, it's enough to match your preference, like "I want the conclusion briefly" or "I want a polite report."

15. Don't cram everything into Custom Instructions

Custom Instructions are convenient. But if you cram in company rules, this month's numbers, product info, SNS rules, customer info, and writing rules, it actually becomes harder to understand.

Only put things you use commonly. Put information for each project into that project's folder or materials. Create a place for information too. This is important.

16. Decide "how much it can do on its own"

I think this is the most important thing in the Astra era. For example:

Reading materials
Fixing text
Organizing tables

For these, let it proceed as is. On the other hand:

Sending emails
Deleting files
Purchasing products
Incurring charges
Changing permissions

For these, have it check with you once. In other words, delegate reversible work. Check irreversible work. Just remember this.

17. Tell it "what constitutes completion"

Just telling the AI to "make a report" doesn't tell it when to end. For example, tell it:

One A4 page
For the boss
Readable in 3 minutes
Write the conclusion at the end
Verify numbers with the original materials

In difficult terms, this is called the Definition of Done. In plain Japanese, "Conditions for Completion" is enough.

18. Have it search for the latest information

This is important. AI looks like it knows everything. But things like today's prices, latest services, laws, news, the president's name, and available plans change.

So when latest information is involved, ask it to "Check the latest information on the Web." Furthermore, add "Prioritize official information." This alone can significantly reduce accidents.

19. Decide where to check

If you ask the AI to "check everything," it might check for longer than necessary. For a report, for example, focus checks on:

Numbers
Names
Dates
URLs

For Excel, check:

Totals
Units
Periods
Formulas

Give check items as part of the job. This is the trick.

20. Don't permit everything from the start

When delegating browser or file operations to the AI, checking every time is tedious. So you'll want to permit everything. But for company use, it's safer to start with "check required." Once you use it and find "it's no problem to delegate this task," then you loosen it.

Here, I investigated one more thing

To write this article, I also investigated "whether the numbers written on the official website and the numbers in my environment are truly the same."

What was interesting there was the context length. Context is, simply put, the amount of material the AI can spread out on its desk at once. The wider the desk, the more materials it can look at while working.

However, when I looked into it, the numbers displayed in the API specs, the Codex catalog, and the actual execution environment were not necessarily the same.

In GPT-5.6 Sol, cases have been confirmed on the official GitHub where a value of 272,000 tokens is used on the Codex side. (GitHub

In other words, just because you read an article saying "this model supports 1 million tokens" doesn't mean you can use 1 million tokens in your Codex. This was the most educational part of this verification.

Level 3: 10 Items to Look at Once You're Used to It

From here, it's a bit advanced. Beginners, please come back when you need this.

21. Separate the AI browser and your everyday browser

In the browser you usually use, you are logged into many things like email, banks, SNS, and company systems. When delegating browser operations to the AI, you don't need to show it everything from the start. Start with a separate browser for the AI. Only use your usual browser when necessary. This is safer.

22. Narrow down the sites it can access

It's convenient when the AI can operate the Web. But it doesn't need to be able to go everywhere. Only permit sites you use daily. Restrict unnecessary sites. In company terms, just because you have an ID card doesn't mean you can enter every room in the company. It's the same.

23. Connect only necessary external services

Google Drive. Email. Calendar. Internal systems. Connecting them to the AI makes things convenient. But the more you connect, the wider the range the AI can access. At first, only the services you truly use. This is enough.

24. Don't increase Skills too much

Skills are, simply put, a mechanism to have the AI learn how to do specific jobs. For example, "Skill to research competitor products" or "Skill to check SNS posts." Having 100 skills doesn't make it 100 times more excellent. Rather, it's more important to know when to use which skill and what it does.

25. Review the AGENTS.md you made long ago

This is for people who have been using Codex for a while. Are you adding rules like "Always read everything," "Always check," or "Test everything" every time the old AI failed? Those rules might not be necessary for the current Astra. When you change to a new model, see if you can delete old settings instead of just adding new ones. This is also important.

26. Don't think you can "put everything in" just because it's 1 million tokens

Even if a huge context is available, you don't need to give all the materials at once. Imagine a subordinate who put 10 years' worth of meeting materials on the desk at once. Even if the desk is wide, searching is hard. Being able to reach the necessary materials quickly is more important.

27. Be careful when fixing the default model in Codex

When you use Codex heavily, you might fix the model you use every time in a configuration file. It's convenient, but just because you wrote it in the config file doesn't mean that value is truly effective. This was something I was concerned about in my measurements this time.

[Deputy Manager Taro's Measurement Note ②] "Being able to write it" and "It being effective" were different

To investigate the values that can be accepted as settings, I tried several candidate values. What concerned me was that even if I put in an obvious test value, there were cases where the command itself ended with exit code 0. In other words, at least in the commands I checked this time, "being able to write the setting" and "that setting being correct" were not the same. The same can be said for office workers. You wrote the setting in ChatGPT. You wrote the setting in config.toml. You wrote the rule in Custom Instructions. Don't be satisfied there; check once if it's truly running with that setting. It's better to think that checking is part of the setting.

28. Keep network and file permissions to a minimum

When you delegate work to the AI, the things it can do increase: it can write files, search the Web, and access external sites. But only permit the range necessary for the job. This is the basic rule. An "AI that can do only what's necessary" can be easier to handle in a company than an "AI that can do everything."

29. Don't use the "strongest" setting every time

When you start using AI, you'll want to use the setting that seems to have the highest performance. But the ability needed for a job like fixing 3 lines of an email is different from a job like analyzing a company's business plan. Light work should be light. Difficult work should be deep. This distinction is important.

30. In a company, decide "who is permitted to do what"

If it's personal, you're done after setting it yourself. In a company, it doesn't work that way. Who uses Astra? What kind of data can they handle? What kind of services can they connect to? Do you permit external transmission? Only after deciding these can you say you've introduced AI as a company.

What surprised me the most this time

While researching the 30 items, I noticed one thing apart from Astra itself. I initially thought I could understand the settings by looking at the official website. However, when I checked the same 5 items in a different environment, quite a few remained "unmeasured" or "unconfirmed."

In one environment, the Codex CLI itself was missing. In another, Codex was there, but the version was old, and Astra was not in the model list. In other words, what's written officially is "what the product can do." On the other hand, how your environment is currently running can only be known by looking at your environment. This was a fairly big difference.

It's not "It's different from the official!"

I'll write this to avoid misunderstanding. Even if multiple numbers like 272,000, 872,000, and 1,050,000 appear for the context length, it doesn't mean one of them is simply a lie. The upper limit available as an API. The upper limit Codex allows in that environment. The effective value on the screen. There's a possibility that what you're looking at is different. In fact, on the OpenAI Codex GitHub, cases have been confirmed where 272,000 tokens are set/reported on the Codex side. So when looking at numbers, don't just look at "how many tokens?" but also "the numbers for which product, on which screen?" This is important.

If you're doing it today, just these 5

30 items is impossible. If you thought that, you're right. I don't intend to check 30 items every morning either. If you're doing it today, just these 5:

1. Check where you can use Astra

2. Separate "work that runs on your PC" and "work that runs in the cloud"

3. If using at work, check Data Controls

4. Decide "work that can proceed on its own" and "work you want to be checked"

5. Write "completion conditions" when asking the AI for a job

This is enough for starters.

Finally. Setting up Astra is like "how to delegate work to a brilliant subordinate."

I thought researching 30 settings for GPT-6 Astra was difficult. But when I replaced it with company work, it suddenly became easier to understand. When you ask a subordinate to do a job, you don't just say "just do your best." You tell them:

What you want them to make
Which materials they should look at
How much they can decide on their own
When you want them to consult you
What constitutes completion

AI is the same. So, when I ask Astra for a job, I decided to look at these four things first:

Purpose: What do I want to complete?

Materials: What can it look at?

Scope of delegation: How much can it decide on its own?

Completion conditions: What needs to be done to finish?

This comes before memorizing difficult prompts. And one more thing I learned from the measurements this time:

"I set it" and "It's running with that setting" are different.

Read the official website. Make the settings. Finally, check once in your own environment. In an era where product updates are fast like GPT-6 Astra, I think including that is what "initial setup" means. As of September 7, 2026, fixes are continuing on the Codex side following the Astra support. In fact, in 0.153.4, a fix to display Astra in the bundled model picker has been released.

So this article isn't the "eternal correct answer" either. The important thing is not to assume that the settings you read yesterday are still correct today. When you put in a new AI, you increase new settings. Before that, try reviewing your current settings once. I started from there too. This way of thinking will be useful when new models come out in the future. Now, let's move on to the intermediate articles and beyond. Thank you for reading to the end.

[Open Chat for office workers who want to reduce overtime with AI has started!]

🎁 Participation Benefits 🎁

AI Work Strategy Starter Kit for Zero Overtime

AI beginners and those who just want to watch are also welcome! First, please feel free to take a look!

👇 Please join from here.

https://app.notion.com/p/10-3b78eeee609c8070997bc1ef986488ef?source=copy_link