If you think of Codex as just a "chat that writes code," you're already a generation behind.
On July 9, 2026, GPT-5.6 was released to the public.
At its center is the flagship model, "Sol."
Its ability to carry out a single request to the end—from complex coding, research, and document creation to browser operations, Computer Use, security, and long-term project execution—has grown significantly.
It recorded an 80 on the Artificial Analysis Coding Agent Index. On Terminal-Bench 2.1, it reached 88.8%, and in Ultra settings, it climbed to 91.9%.
However, there is a change even bigger than the numbers.
It is that Codex has transformed from an "AI that answers questions" into an "AI that assembles and completes work."
Researching.
Planning.
Creating.
Checking.
Distributing tasks to multiple AIs if necessary.
Saving completed procedures and running them automatically from the next time.
You can complete this entire cycle within Codex.
Already, over 5 million people use Codex weekly, and about 20% of them are non-engineers. Moreover, usage among non-engineers is growing more than three times faster than among developers.
In other words, this change is not just for engineers.
Article production, SNS management, competitor research, product planning, document creation, customer support, and web production.
Almost any work done on a computer is a target.
In this article, I will connect all the functions necessary to fully utilize Codex in the GPT-5.6 Sol era, in order of increasing power.
From choosing between Sol, Terra, and Luna to Plan mode, AGENTS.md, config.toml, Skills, Plugins, MCP, Ultra, Subagents, Custom Agents, reviews, and Automations.
This is not a fragmentary introduction of features, but a strategy map for completing your own dedicated work environment.
For those who want to know the big picture of achieving results in a side hustle along with this article 🎁
Currently, on the official LINE,
"Kuroneko-style SNS Side Hustle Complete Strategy: 5 Major Bonus Pack"
is being given away for free 🎁
Since this was originally planned to be released as paid content,
distribution will close once the capacity is reached.
Please receive it along with the article while you can.
▼▼▼
▶︎▶︎▶︎ Receive the 5 Major Bonuses
Now, let's get into the main topic!
The true identity of Codex is not "AI Chat," but an OS that runs work
To understand Codex in the Sol era, you first need to grasp the big picture.
Codex consists of the following 6 layers.
Many people only look at the first layer: "which model is the smartest."
However, the difference in practical work is made from the second layer onwards.
No matter how smart a model is, if the purpose is vague, necessary materials are missing, and completion conditions are not set, what comes back will be a general theory that doesn't reach the goal.
Conversely, if you provide context, rules, tools, roles, and completion conditions, Codex turns into the side that completes the deliverables.
There is no doubt that Sol is strong.
But just choosing Sol does not complete Codex.
You only see its true potential when you connect the performance of the brain to the mechanism of work.
1. Choosing between Sol, Terra, and Luna for work
In GPT-5.6, you can choose from three models depending on your purpose.
Sol is the "Command Center" that thinks through to completion
Sol is the flagship model of GPT-5.6. It is suitable for work where the answer is not decided from the beginning.
- Reading multiple materials to decide on a strategy
- Understanding a large codebase to add features
- Proceeding from research to structure, production, and verification all at once
- Completing deliverables across browsers and apps
- Maintaining consistency in long projects
- Commanding multiple subagents
Sol is for work where you want the AI to think not just about "what to make," but also "how to proceed to achieve the goal."
Terra is the "Practitioner" balancing speed and quality
Terra is a model that balances capability and cost. It is suitable for work that requires judgment but doesn't need the deep, continuous thinking of Sol, such as daily research, summarization, file organization, drafting, code correction, and checking multiple materials.
When running multiple subagents, assigning Terra to research or exploration roles increases efficiency.
Luna is the "Worker" in charge of volume and speed
Luna is the fastest and lowest-cost model.
- Categorizing large numbers of files
- Checking for notation inconsistencies
- Primary screening
- Converting boilerplate text
- Generating large numbers of candidates
- Formatting into a fixed style
It is suitable for running these light tasks at high frequency. Instead of leaving the final judgment to Luna, let it handle gathering, organizing, and narrowing down candidates, then return to Sol at the end. This division of labor is powerful.
If you're unsure, start with this combination
You don't need to run Sol at maximum settings every time. Sol for the command center, Terra for research, and Luna for routine processing. Just like a human team, you allocate brains based on the weight of the work.
Ultra is not just a "deep thinking setting"
Ultra is a setting that uses the highest level of reasoning for the corresponding model. Even more important is its ability to actively distribute suitable tasks to multiple subagents.
In normal settings, you can parallelize by explicitly saying "divide this among three people." In Ultra, Codex judges that "dividing this work will be faster and improve quality," and can decompose it into research, production, verification, etc. In short, Ultra is not just a high-intelligence mode; it is an automatic AI team formation mode.
Fast mode increases speed without changing the model
Codex also has a Fast mode. In exchange for increasing the speed of the corresponding model by about 1.5 times, it consumes more credits in GPT-5.6 than usual. You can switch it in the CLI with the following commands:
/fast on
/fast off
/fast status
This feature is for corrections with short deadlines or focused work where you want to reduce waiting time. Fast mode is different from "dropping to a lighter model." Use it when you want to increase speed while maintaining Sol's capabilities.
2. Choosing between App, CLI, IDE, and Cloud
Codex's strengths change depending on where you use it.
The ChatGPT Desktop App is the "Command Room"
If you want to plan while looking at multiple files, proceed with tasks, and handle images, documents, tables, browsers, and external tools, the desktop app is the center. You can manage Codex's progress, diffs, subagents, Skills, Plugins, and Scheduled tasks all together. For non-engineers looking to integrate Codex into their work, starting with the app is the shortest path.
CLI is the "Executioner in the Terminal"
The CLI is good at directly handling local files and code, executing commands, testing, Git operations, and non-interactive automatic processing. By using codex exec instead of just the interactive codex, you can run it from scripts or CI. The value of the CLI increases for those who want to run fixed processes in the same way every time.
IDE Extension is the "Assistant next to the code"
If you want to proceed with corrections, explanations, and reviews while looking at the code you have open in VS Code or similar, use the IDE extension. It's easy to give detailed instructions while switching target files, making the back-and-forth during implementation the shortest.
Cloud is the "Outsourcee that frees up your computer"
Cloud is suitable when you want to leave time-consuming tasks to a separate environment. You can proceed with other tasks in parallel without stopping your local work.
The thinking is simple:
- Daily command: Desktop App
- Commands and automatic processing: CLI
- Close-contact work for code implementation: IDE
- Long-duration separate tasks: Cloud
You don't need to consolidate everything into one. Use the same Codex from the entrance that fits the work.
3. Provide instructions with four elements: Purpose, Context, Constraints, and Completion Conditions
GPT-5.6 Sol can move quite well even with short instructions. Still, for important work, providing these four elements makes it overwhelmingly more stable.
Purpose
It's not about what to make, but what you want to achieve. Instead of "write an article," say "complete an article that allows readers who only use Codex for one-off chats to create their own dedicated AI work environment."
Context
Which files, materials, examples, and past decisions should it look at? The strength of being able to give folders to Codex lies here. Instead of rewriting explanations in the chat box every time, let it read the correct materials.
Constraints
Conditions to be followed. Include word count, tone, files that must not be touched, technologies to use, target audience, primary information to refer to, prohibited expressions, etc.
Completion Conditions
What needs to happen for the work to be finished? Instead of "finish after writing the text," decide "complete after finishing fact-checking, link checking, word count checking, readability checking, and saving to the specified folder."
Summarizing these four results in the following format:
ーーーーーーーーーーーー
【Basic Prompt for Handing Work to Codex】
Purpose:
[What you want to achieve with this work]
Context:
[Files, folders, reference materials, past decisions to read]
Constraints:
[Rules to follow, scope of change, target audience, format]
Completion Conditions:
[What to check and what state to be in for completion]
Please proceed with the necessary research and work yourself, and execute until the completion conditions are met. Ask questions only where judgment is required, and otherwise judge and proceed rationally.
ーーーーーーーーーーーー
Sol moves more strongly when the purpose and completion conditions are clear than when you list 30 detailed steps. If you decide all the steps, Codex can only do the work it was instructed to do. Clarify the goal and leave room for the process. That is the instruction for an agent.
Turn information given to Codex into files, not chat
The longer you use Codex, the more file design matters over chat skill. If you proceed only with conversation, important decisions, reference materials, deliverables, and next tasks get mixed in the same place. You have to re-explain every time you start a new chat, and a different judgment comes back from last time. To escape this state, decide "what needs to be read to resume work" for each project.
The minimum configuration is the following four:
Project/
├── Context.md # Purpose, target, premises that rarely change
├── Project.md # Current issues, decisions, next tasks
├── Materials/ # Reference materials, source data, competitor info
└── Outputs/ # Completed deliverables
Place "premises that rarely change" in Context.md
Place information that is needed every time, such as project purpose, target audience, judgment criteria, and conditions to be followed.
Place "what is being done now" in Project.md
Update current issues, plans under consideration, decisions, and the next move. Even if the chat changes, you can start from where you left off by reading this file.
Place "evidence" in Materials
Summarize materials for creating deliverables, such as reference articles, competitor research, images, minutes, data, and specifications.
Place "final versions" in Outputs
By separating drafts and final versions, it becomes harder for Codex to mistake an old draft for the master copy.
Once you've created this structure, write the reading order in AGENTS.md. Then, the next request can be short:
Follow the AGENTS.md of this project and read Context.md and Project.md. Proceed from the current point until the completion conditions are met.
Chat is the place for instructions and judgment. Files are the place for memory and deliverables. When this division of roles is achieved, Codex becomes not a one-time conversation partner, but a person in charge of continuously advancing the project.
4. Start with Plan mode for vague work
You have something you want to do, but you don't know what to make or where to start. If you jump straight into implementation or production in this state, the premises will shift midway. That's where Plan mode comes in.
In Plan mode, Codex first researches files and the situation, asks necessary questions, and creates a blueprint before execution. You can switch to it with /plan in the CLI or Shift+Tab in the app.
Plan mode is strong for the following work:
- New projects where requirements are still vague
- Changes spanning multiple files
- Renovations where you don't want to break existing mechanisms
- Tool introduction with many options
- Process design for long-term projects
- Article production decided from the direction of readers or products
Usage is not difficult.
ーーーーーーーーーーーー
【Prompt for Plan mode】
Do not execute this request immediately; first, research the current situation.
- Gather information necessary to achieve the purpose
- Separate unclear points from important judgments
- Plan execution steps, change targets, and verification methods
- Ask only the things I need to decide
Once the plan is solid, present it in an executable order.
Purpose: [What you want to achieve]
ーーーーーーーーーーーー
The value of Plan mode is not being cautious; it's eliminating rework. It's faster to spend the first 15 minutes creating a correct design than to start making it in 10 minutes and redo it 3 hours later. The larger the work, the wider this gap becomes.
5. Eliminate "repeated explanations" with AGENTS.md
The first asset someone starting with Codex should create is AGENTS.md. AGENTS.md is the rulebook Codex reads before starting work. You can fix things you want it to follow every time in a file instead of chat.
For example, the following content:
- Order of files to read first
- Project purpose
- Important folders
- Rules for writing and design
- Test and confirmation commands
- Scope that must not be changed
- Definition of completion
- Reporting method to the user
Separate Global and Project
Place common personal rules in ~/.codex/AGENTS.md. Place rules for a specific project in AGENTS.md directly under the project. If a specific folder needs separate rules, you can add an AGENTS.md inside that folder. Codex reads from the top-level rules and prioritizes files closer to the workspace. In other words, you can separate general rules from on-site rules.
The first AGENTS.md is enough with this
AGENTS.md
Purpose
- What to achieve in this project
Read First
- Context.md
- Project.md
- Specifications of the target function
Work Rules
- Do not delete existing data
- Prioritize existing design patterns
- Do not change unrelated files
Completion Conditions
- Necessary implementation or deliverables are completed
- Testing and display confirmation are finished
- Report change details and confirmation results
You don't need to create an encyclopedia from the start. When Codex makes the same mistake, add the rule that caused it. If you give the same explanation twice, it's a problem with the mechanism, not the conversation. Don't just fix it on the spot; change it so it doesn't happen next time. AGENTS.md is the place to grow Codex.
6. Set Codex's initial state with config.toml
If AGENTS.md is the "work rules," config.toml is the "Codex main body settings." It mainly manages the following items:
- Model to use
- Reasoning effort
- Permissions and approval methods
- Sandbox
- MCP servers
- Subagent settings
- Feature flags
- Profiles
Place personal settings in ~/.codex/config.toml. Place project-specific settings in .codex/config.toml. CLI, IDE extensions, and desktop apps share this setting layer.
In a minimum configuration, it looks like this:
model = "gpt-5.6"
model_reasoning_effort = "high"
approval_policy = "on-request"
[agents]
max_threads = 6
max_depth = 1
max_threads is the number of agent threads that can be opened simultaneously, and max_depth is the depth to which subagents can further branch down. The current standard is a maximum of 6 threads and a depth of 1. You don't need to recursively increase a large number of agents from the start. It's strong enough for the main to hand work to multiple specialists and collect the results.
Dividing the roles of settings prevents confusion:
- How to behave: AGENTS.md
- Which model, permission, and connection to use: config.toml
- How to proceed with work: Skills
- What to do with external services: MCP / Plugins
7. Turn "successful procedures" into abilities with Skills
Are you explaining work you do every week from scratch every time? Article production, competitor research, meeting summarization, releases, reviews, invoice processing, report creation. If you repeat the same process, the next thing to create is not a long prompt, but a Skill.
A Skill is a work-specific ability that can be added to Codex. Basically, you write the following in SKILL.md:
- When to use it
- What input to receive
- What to read
- In what order to proceed
- Which tools to use
- What to check for completion
If necessary, reference materials, templates, scripts, and image assets can be placed in the same folder.
Skills read the full text only when necessary
Codex does not load all Skill texts from the beginning. It first looks at the name and description and opens only the Skill that fits the current request. This is "Progressive Disclosure." You can call only the necessary abilities without stuffing a large number of procedures into the context every time.
Can be used explicitly or automatically
When explicit, specify $skill-name in the prompt. If the description and request content match, Codex can also choose automatically. That's why the description is more important than the Skill name. What kind of Skill is it, when to use it, and when not to use it? If this is clear, false activations will decrease.
Work that should be made into a Skill
If two or more of the following apply, it's time to turn it into a Skill:
- You've done the same process 3 or more times
- The materials referred to are the same every time
- Quality drops if the order is wrong
- There are check items that must always be passed
- Coordination with a specific tool is required
- You want to reuse it with other people or in other projects
Just saving a prompt that worked once does not increase reproducibility. It becomes an ability only after fixing the input, process, judgment criteria, and verification.
If you use macOS where Computer Use is available, you can create Skills from demonstrations
For operations that are difficult to explain in text, you can use Record & Replay. If you actually show the operation on a Mac, Codex analyzes the procedure and creates a draft of the Skill. This feature is compatible with "work that is faster to show than explain," such as expense reimbursement, downloading routine reports, posting videos, and filling out fixed forms.
8. Plugins bundle "Abilities, Connections, and Tools"
If a Skill is a work procedure, a Plugin is a package that distributes multiple abilities and connections. A Plugin can bundle the following elements:
- Skills
- Connectors such as Gmail and Google Drive
- MCP servers
- Hooks
- Browser features
- Scheduled task templates
Before making a Skill yourself, if there is a Plugin that fits your purpose, it's faster to use the existing one first. For example, by adding Plugins for GitHub, Gmail, Google Drive, and Slack, Codex's workspace expands beyond local folders.
Difference between Skill and Plugin
Plugins are available from the Plugin browser in the desktop app and CLI. In the CLI, open it with /plugins. After installation, starting a new chat or session makes the added Skills and tools available.
9. Give Codex "hands and feet for external services" with MCP
No matter how strong Codex is, it cannot touch the latest data or private information of services it is not connected to. You want it to read materials from Google Drive. You want to check GitHub Issues. You want to see Figma designs. You want to get information from Notion or internal systems. You want to operate a browser. That's where MCP comes in.
MCP is a common standard for connecting Codex with external tools and information. MCP servers mainly provide three things:
- Tools: Operations such as search, create, update, and send
- Resources: Reading documents, data, specifications, etc.
- Prompts: Reusable prompts for that service
Adding MCP changes the form of the request
Before connection, the user gathers information and pastes it into Codex. After connection, Codex itself can get necessary information, create deliverables, and reflect them in necessary places. For example, work like this:
- Gather meeting materials from Google Drive and summarize decisions
- Check GitHub PRs and Issues and implement fixes
- Look at Figma to reproduce screens and check display in a browser
- Extract emails needing replies from Gmail and create drafts
- Create implementation plans from Notion specifications
Combine Skill and MCP
MCP alone just increases tools. Write the order of what to do in a Skill. "Every Monday, read values from Drive, compare with last week, check for outliers, and create a report." In this case, the hand that takes information from Drive is MCP, and the weekly work procedure is the Skill. Separate tools and procedures. This idea stabilizes Codex.
10. Create a "One-Person AI Team" with Ultra and Subagents
The highlight of the Sol era is not making a single AI even smarter. It is making multiple AIs work simultaneously. Codex can divide work into subagents, proceed in parallel, and finally, the main agent can integrate the results.
Letting one person do everything pollutes the context
If you keep putting long research logs, test results, errors, candidate plans, and rejected plans into one chat, important purposes and judgments get buried. This is Context pollution. Furthermore, as unnecessary information continues to increase, judgment accuracy drops in the latter half of a long conversation. So, offload heavy intermediate work to separate agents.
- Main: Purpose, judgment, integration, final version
- Researcher: Materials, competitors, facts, numbers
- Creator: First draft, implementation, candidate creation
- Verifier: Mistakes, omissions, shifts, tests
Return only organized conclusions to the main, not the long work logs of each person in charge.
Built-in 3 roles
Codex has three basic agents:
default: General purposeworker: Proceeds with implementation and correctionsexplorer: Reads and researches code and materials
These three are enough at first. If you want to further fix roles, you can create Custom Agents. Place TOML files in ~/.codex/agents/ for personal use and .codex/agents/ for project use. In addition to name, description, and assignment instructions, you can also change the model, Reasoning, Sandbox, MCP, and Skills for each role.
First 4-person team to create
ーーーーーーーーーーーー
【For Copy-Paste: Sol Command Center + 3-Person AI Team】
Proceed with this work using a main agent and three subagents.
Main Agent:
Manage purpose and completion conditions, and create the final version from the results of each person in charge.
Research Lead:
Gather necessary primary information, examples, numbers, and premises, and return them with evidence.
Production Lead:
Create the first draft of the deliverable based on research results and purpose.
Verification Lead:
Check facts, omissions, quality, readability, and shifts from the purpose.
Execute work that can be proceeded with independently in parallel. Wait for all leads to complete, and let the main agent integrate the results.
Final Output:
- Final version
- Basis for adoption
- Points fixed in verification
- Remaining judgments
Purpose: [Purpose here]
Completion Conditions: [Completion conditions here]
ーーーーーーーーーーーー
Parallelize "independent work"
Increasing subagents doesn't make everything faster. What's strong is work that can be done at the same time.
- Research of multiple materials
- Separate perspective reviews for security, quality, and readability
- Classification of large numbers of files
- Creation of multiple plans
- Testing and log analysis
On the other hand, if multiple people rewrite the same file at the same time, they will conflict. Consolidate the writing role to one person, and parallelize the reading, researching, and verifying roles. This is the first correct answer.
11. Separate the Production Lead and Verification Lead
Just asking Codex to "make it, then check if there are any problems" will repeat production and confirmation from the same perspective. If you want to improve quality, separate roles from the beginning.
For code:
- Implementation lead
- Test lead
- Security lead
- Maintainability review lead
For articles:
- Writing lead
- Fact-checking lead
- Beginner's perspective lead
- Title consistency check lead
For materials:
- Structure lead
- Numerical confirmation lead
- Design confirmation lead
- Decision-maker perspective lead
Even for the same deliverable, the points raised change when the role looking at it changes. Codex also has /review. You can proceed with reviews after implementation targeting uncommitted changes, specific commits, diffs with the base branch, etc. However, calling the review function is not enough. Decide what to find as a problem.
ーーーーーーーーーーーー
【For Copy-Paste: Pre-completion Review】
Check this deliverable as a reviewer separate from the creator.
Priority:
- Defects that prevent achieving the purpose
- Errors in facts, numbers, or specifications
- Missing premises or steps
- Places where the user will get lost
- Readability, maintainability, expression
List problems in order of importance and show the relevant parts and suggested fixes. If there are no problems, briefly show the scope confirmed and remaining risks.
Completion Conditions: [Completion conditions here]
ーーーーーーーーーーーー
Don't say "make it look good"; give passing conditions. Verification is also part of the work.
12. Design the "scope to entrust" with Permissions and Sandbox
Codex can read and write files, execute commands, and operate external services. That's why permission design is as important as the model's intelligence. There are three basic ideas:
- Read only: Just reading
- Workspace write: Can change within the workspace folder
- Full access: Can access a wide range
Base daily production and implementation on Workspace write. Add necessary ranges only when external networks or separate folders are needed. And leave confirmation for operations that are hard to undo, such as deletion, sending, publishing, payment, and changing external services. This is not to make Codex weak. It is the foundation for entrusting large tasks with peace of mind. If permissions are vague, Codex will stop at necessary operations or, conversely, have too wide a range. Deciding "how far it can proceed automatically" first reduces the number of confirmations during work.
13. Automate repetitive work with Automations
Once work is successfully done once, automate it next. Using Codex's Scheduled tasks, you can run work at fixed times, regular intervals, after events, or under monitoring conditions. For example, usage like this:
- Every morning, gather the latest news in the AI industry
- Every week, check new articles from competitors
- Every night, review project changes
- Regularly check PR status and respond to new points raised
- Create a report at the beginning of the month
- Track until long-duration processing is finished in the same chat
Distinguish between one-off tasks and continuous chats
If you want independent results every time, use a Standalone Scheduled task. If you want to carry over previous conversations and follow the same work, create a schedule within an existing chat.
Keep the app running for local work
Scheduled tasks that handle local projects in the desktop app require the computer and app to be running. In Git projects, you can choose whether to use the current workspace folder directly or separate it with another Worktree. If there's a possibility that a periodic task will touch files you're currently working on, it's easier to handle by separating with a Worktree.
Automation is done "after manual success"
Don't run it every day suddenly; first complete it once in a normal chat. Next, make it a Skill. Finally, put it on a Scheduled task. Manual success → Skillization → Automation. In this order, you won't mass-produce wrong work every day.
14. Start from this configuration based on purpose
You don't need to use all features. Build from the layers necessary for your work.
AI Beginners / Employees
- Desktop App
- GPT-5.6 Sol or Terra
- Purpose, Context, Constraints, Completion Conditions
- Plan mode
- AGENTS.md for the project
The first goal is to hand one folder to Codex and proceed from planning to completion.
Article, SNS, Content Production
- Use Sol for structure and final editing
- Use Terra for research
- Save production rules in AGENTS.md
- Skillize article production and post creation
- Connect Web search and Drive with MCP
- Subagentize the fact-checking lead
- Make topic research a Scheduled task
With this configuration, it connects not just making text, but planning, research, production, confirmation, and next improvement.
Sole Proprietors / One-Person Companies
- Separate folders and master copies for each business task
- Place general rules in AGENTS.md
- Turn routine tasks into Skills
- Connect Gmail, Drive, GitHub, etc., with Plugins/MCP
- Create Custom Agents for research, production, and verification
- Distribute multiple projects with Ultra
- Move stable tasks to Automations
The goal is not to be a person who asks AI questions, but a person who distributes work to AI and only judges the results.
Developers / Production Teams
- CLI or IDE Extension
- AGENTS.md directly under the repository
.codex/config.toml- Fix lint, test, and build to completion conditions
- Distribute implementation, testing, and review among subagents
/reviewand GitHub integration- Move PR monitoring and periodic reviews to Scheduled tasks
Don't stop at code generation; complete the cycle through testing, diff confirmation, review, and PR response.
15. 7 Common Traits of People Who Fail with Codex
- Stuffing everything into one chat If you continue research, production, correction, and separate projects in the same chat, the purpose gets buried. Separate projects and offload heavy intermediate work to subagents.
- Giving the same explanation every time Move repeated premises to AGENTS.md and repeated processes to Skills. Don't shorten the conversation; turn explanations into assets.
- No definition of "completion" If you finish just by making it, unverified deliverables will increase. Include tests, confirmation items, save locations, and formats in completion conditions.
- Processing everything with Sol Ultra Separate heavy work and light work. Final judgment is Sol, daily work is Terra, and volume processing is Luna. This division organizes speed and usage.
- Just increasing tools Even if you put in a large number of MCPs or Plugins, they won't work if the process to use them is not decided. Decide the work first and connect only necessary tools.
- Letting the Production Lead grade themselves Separate the role that makes and the role that confirms. For important deliverables, bring in the eyes of another agent.
- Automating before success If you put a procedure you don't know will work into a Scheduled task, confirmation work will increase. Complete manually, solidify into a Skill, and finally automate.
Complete your Sol-era Codex environment in 7 days
You don't need to learn everything today. Build up one layer of work per day.
Day 1: Entrust one task to the end
Open the target folder and provide Purpose, Context, Constraints, and Completion Conditions. Request a deliverable, not a question.
Day 2: Let Plan mode design
Choose one vague task and entrust it with research, questions, and planning. Get a feel for eliminating rework before execution.
Day 3: Create AGENTS.md
Write just 5 things you explain every time. Including reading order, rules to follow, and completion conditions is enough.
Day 4: Turn repetitive work into a Skill
Choose work you do at least once a week and fix the input, process, and confirmation method.
Day 5: Connect one external service
Connect the one you use most frequently, such as Drive, GitHub, Gmail, or a browser, via Plugin or MCP.
Day 6: Run 3 subagents
Divide into research, production, and verification, and finally integrate with Sol. You'll see the difference from when one person proceeds in order for the same work.
Day 7: Add review and automation
Add review to completion conditions and move one stable task to a Scheduled task.
By this point, Codex is not a one-off chat. It becomes a work environment that reads your rules, uses necessary tools, divides work among multiple leads, and confirms until completion.
Quick Reference Table to Look at Last
What is needed in the Sol era is not prompt skill, but work design skill
With GPT-5.6 Sol, Codex has become even smarter. But the truly big change is not the numbers on the performance chart. It is that AI can now think about necessary work from a purpose, read materials, use tools, distribute to multiple AIs, and proceed to completion without humans instructing each step one by one.
From now on, the difference will be made not by those who know magic prompts, but by those who can prepare the correct context. Those who can turn repeated judgments into rules. Those who can save successful processes into Skills. Those who can connect necessary tools with MCP. Those who can divide work among multiple AIs and manage with completion conditions.
In other words, not people who use AI, but people who create an environment where AI can work.
The era of just opening Codex and throwing out questions on the spot is over.
Create a folder. Place master copies. Decide rules with AGENTS.md. Let Skills learn work. Give hands and feet with MCP. Move the team with Ultra. Confirm completion with Review. Make it automatic from next time with Automation.
For those who create this cycle, Codex ceases to be "convenient AI." It becomes a team that works longer than you, reads more information than you, and proceeds with work according to your rules. That is Codex in the GPT-5.6 Sol era.
For those who want to know the big picture of achieving results in a side hustle along with this article 🎁
Currently, on the official LINE,
"Kuroneko-style SNS Side Hustle Complete Strategy: 5 Major Bonus Pack"
is being given away for free 🎁
Since this was originally planned to be released as paid content,
distribution will close once the capacity is reached.
Please receive it along with the article while you can.
▼▼▼
▶︎▶︎▶︎ Receive the 5 Major Bonuses
Now, let's get into the main topic!
