How to Apply Codex to Office and Knowledge Scenarios

Foreword

Many people’s first impression of Codex is that it’s just an AI tool for writing code, irrelevant to them.

That judgment is wrong.

Imagine a Monday morning: You receive an email. Your boss wants you to produce a product launch plan. You open Codex, drop the requirements in, close your laptop, and head out. On your commute, your phone pings with a notification—Codex has already read through the relevant Feishu/Slack discussions, dug up customer research from Google Drive, checked last quarter’s data, and started drafting the plan in Notion. It asks you a question about a timeline. You tap to confirm. By the time you get to the office, a draft is waiting for you.

Codex’s core capabilities were originally designed for writing code, but this same set of skills—pulling information from multiple sources, calling various tools, executing complex tasks step by step—works just as well for office work.

This guide teaches you how to use Codex for your daily work. From setting up your environment, to running actual workflows, to turning repetitive tasks into an increasingly smooth system.

You don’t need to learn every feature at once. First, know what’s available, then gradually incorporate it at your own pace. This article is long; consider bookmarking it.

Part 1: What is Codex

One-Sentence Explanation

Codex is an AI workspace. You tell it what to do, and it figures out how, finds the tools, does the work, and finally hands you the result for review.

Things it can do include:

• Read and write files in your project folder
• Connect to external tools like Feishu, Slack, Gmail, Notion, Google Drive, etc.
• Automatically run multi-step processes
• Write code and scripts when needed
• Output various formats: documents, spreadsheets, PPTs, PDFs, web pages, etc.
• Handle multiple things at once
• Remotely start and monitor tasks from your phone (your computer needs to stay on)
• Turn frequently done tasks into automated workflows
• Share your work results as a publishable webpage

Five Core Concepts

There are five major concepts in Codex. Just remember one sentence for now: Project is the container, conversations and goals run inside it, plugins extend capabilities, sites output results.

Project = A workspace. Create one for “product launch” and another for “weekly report.” All conversations within a project share the same set of files and rules. Store important things (decisions, status, key links) in project files—don’t rely on Codex’s chat history to remember.

Thread = A specific conversation within a project. Open one thread per task. If the task changes or the conversation gets messy, start a new thread. Note: Different threads don’t automatically share chat history, but they can pass information through shared files—for example, thread A writes results into a file, and thread B reads that file.

Goal = A long-term task set with the /goal command. Suitable for work where you know the final outcome but need many intermediate steps. You tell Codex what counts as “done,” and it keeps advancing, notifying you when complete or stuck.

Plugin = A packaged toolkit. For example, someone might make a “write weekly report” plugin that you can install. Before building from scratch, browse the plugin directory to see if something already exists.

Site = Turn your work results into a webpage. Suitable for scenarios where others need to see it and interact with it (e.g., a project dashboard). If someone only needs to read it, a document is sufficient; no need for a site.

What Codex Cannot Do

Codex requires human oversight. It cannot make judgments for you, verify facts for you, or take responsibility for you. Do not let it handle the following:

• When it doesn’t have access to the data source
• When the quality depends entirely on subjective judgment
• When the consequences of a mistake are severe

What Tasks Are Suitable for Codex

A task is suitable if it meets two or more of these criteria:

• Needs to pull and aggregate information from several places
• Has steps you repeat frequently
• Can be checked against a standard after completion
• Ultimately produces an output (document, report, plan, etc.)
• Is so annoying you’ve been putting it off

There are two modes for using Codex—this distinction runs throughout the article and is important:

Fully Hand Over to Codex: The task process is clear, low risk, and the result is easy to check. You clearly state requirements, Codex does it all, and you review the final output.

You and Codex Work Together: The task is vague, requires your judgment, or needs iterative adjustments to get right. You advance together, adjusting direction as you go.

In short: if a checklist can handle it, hand it over to Codex; if you have to rethink the process each time, work on it together.

Part 2: Setting Up Your Work Environment

Five-Step Work Method

All good Codex workflows run on the same loop. Each time you run it, you get better and faster:

Connect Tools → Write Rules → Work → Check → Consolidate

• Connect Tools: Integrate the systems you use daily—Feishu, Gmail, Slack, Notion, Google Drive, calendar, data analytics tools, etc. Without connections, Codex can only see files in the project folder and content you manually paste in. PPT, Word, and Excel don’t need additional integration—Codex handles them by default.
• Write Rules: Write your goals, preferences, project details, key links, and check standards into files that Codex can read. The value: after doing it once, Codex won’t need you to re-explain the context every time.
• Work: Decide whether to hand the task fully to Codex or work together. Either way, first clarify: what materials to use, what final format is needed, and what counts as “done.”
• Check: View the results where they will actually be used. If Codex sent a Feishu message, check it in Feishu. If it wrote a document, check it in your usual document tool. Something that looks fine in Codex might be completely wrong in the real context.
• Consolidate: Keep the good stuff. Save effective prompts, document useful processes, add pitfalls to your checklist. Make each iteration faster than the last.

Connecting Your Tools

Connect the tools you use daily to Codex. Note: Connecting a tool does not give Codex unlimited permissions—start with the minimum permissions it needs, and require it to ask you before modifying anything.

There are several ways Codex interacts with external tools. The principle: prioritize the most direct path.

After connecting your tools, you can let Codex help you discover worthwhile workflows to build—expect pleasant surprises:

Connect these tools I use at work: [list your tools]. Look at my work habits in these tools and recommend three workflows to build first. For each, explain: where to pull data from, what the final output is, how often to run it, and what I need to approve.

Setting Up Your Workspace

Before you start working, set up your workspace. Skipping this step will likely cause you to get stuck.

A workspace is a folder containing all the “background information” Codex needs to work—instructions, rules, process documents, common materials, checklists. Think of it as an “onboarding manual”—except the reader is Codex.

Don’t design the file structure from scratch yourself. Let Codex interview you, then have it design it for you:

I want you to help me create a workspace. Ask me one question at a time about my role, responsibilities, current projects, frequently repeated tasks, tools used, collaborators, what outputs look like, my work habits, and anything you must never do without my permission. After finishing the questions, help me design a folder structure and explain what each file is for. Don’t move anything until I confirm.

Reference Folder Structure

No need to follow this strictly. In the early stages, just get started. You can standardize later as you become more proficient. A clear folder structure helps both you and the AI understand the current project status.

What to Put in the Three Key Files

context.md (Who You Are): Your role and responsibilities, current projects, daily tools, main collaborators and teams.

preferences.md (Your Preferences): Writing style and tone, things you need to review before sending, things Codex can just notify you about after doing.

rules.md (Red Line Rules): Things you absolutely must not do without your approval—send messages, publish content, delete things, modify official materials, transfer money. Things you can do without asking—draft, summarize, research, outline, organize materials. Also, any special requirements in your work (e.g., client confidentiality, brand guidelines).

About AGENTS.md

Codex reads the AGENTS.md file before starting any work. Keep a short one in the Codex root directory: explain what this space is for, which files are most important, and which rules apply globally. Put detailed instructions for specific projects in the corresponding subfolders. If you can’t find this file, ask Codex in the chat for the file path, or ask Codex to write it for you. AGENTS.md must be kept brief.

Part 3: Five Levels of Proficiency

Mastering Codex isn’t achieved in a day; it’s a step-by-step process. Each level requires different usage and mindset. Moving too fast leads to frustration—either you don’t trust it yet, or you haven’t prepared what’s needed.

These five levels are steps in complexity, not ranks of ability. Don’t deliberately aim for high-complexity forms.

Level 1: Single Tasks – Start with Simple Things

Treat Codex as a very capable assistant. You give it one task; it does it; you check it. That’s it. No automation, no fancy stuff. You review every result.

Good things to try first:

• Summarize meeting notes, extracting decisions, action items, and issues
• Structure scattered notes into a clear outline
• Pull a research summary from a bunch of links and documents
• Edit your draft according to your style guide
• Create a checklist for a document

Prompt template:

Using these [documents/links/notes], help me create a [specific deliverable]. Accuracy is more important than polish. Label every fact with its source link. Mark anything you’re unsure about. Finally, list three questions I should verify.

A good habit: Before starting to edit what Codex gives you, ask it: “What assumptions did you make while doing this? Where are you least confident?”—Expose issues early, so you don’t waste time correcting the wrong direction.

When to move to Level 2: You start thinking, “If only it could remember what I said last time.”

Level 2: Pulling Information from Multiple Sources – Cross-System Integration

Treat Codex as an analytical assistant that can look at several systems at once. Feishu messages and documents, Notion docs, local documents, Gmail emails—information that would take you half a day to gather can be handled by Codex in one go.

Typical scenarios:

• Combine meeting notes, Feishu messages, and client notes to write a product launch plan
• Pull data from analytics backend, revenue data, and support volume to produce a weekly report
• Synthesize Feishu messages and documents into a weekly report

Prompt template:

I need a [specific deliverable]. Use the following information: [Tool 1]: look for [what info] inside it; [Tool 2]: look for [what info] inside it; [Tool 3]: look for [what info] inside it. Final format: [your desired structure]. Before you start, tell me your plan: what you’ll look up, what you’ll produce, what might be missing, and how you’ll check before finishing. For actions like sending messages or modifying files, ask me first.

Note: Pulling data from multiple systems is error-prone. Data might be outdated, definitions may differ across systems, or data might not match. For numbers that affect business decisions, always go back to the original data source to verify each one.

When to move to Level 3: You find yourself manually running the same workflow every week and start wondering, “Can this be automated?”

Level 3: Scheduled Automation – Free Yourself from Repetitive Work

At this level, you can automate work that is “pretty much the same every time.”

Core judgment: If a checklist can handle it, hand it over to Codex to run automatically; if you have to rethink the process each time, work on it together.

Best “computer chores” for automation:

• Every day before leaving work, scan unread Feishu messages and emails, draft replies or even auto-reply
• Every week, automatically pull data from various tools to generate a metrics brief
• Every time a meeting ends, automatically organize notes and list action items
• Automatically package your draft content into an editor review format

Before setting up automation, think through this table:

Field	Value
Process name:
How often to run:
Where to pull data:
What is the final output:
What Codex can do on its own:
What must ask you first:
How to verify results are correct:
Where to put the results:
When to disable or update this process:

When checking, remember to look where it will actually be used. A Feishu message drafted by Codex—check it in Feishu. An email it wrote—check it in Gmail. Something that looks fine in Codex might feel completely different in the real context.

When to move to Level 4: You find that just writing prompts isn’t enough—some tasks are too complex and need a small script or tool to run reliably.

Level 4: Let Codex Build Small Tools for You

Sometimes Codex’s best output isn’t a document, but a small tool—a script, a simple dashboard, a mini program that automates a specific task.

You don’t need to be a programmer. The approach is:

• First, do it manually in Codex to confirm the result is what you want
• Ask it: “Which steps in this process could be turned into a small tool for automation?”
• Let Codex build a prototype and explain what it does in plain language
• Test it with your real data and see if the result matches the manual method
• Keep what works, discard what doesn’t

You don’t need to understand every line of code, but you need to be able to articulate three things: what is the input, what is the output, and which step needs a human to look at it. If you can’t say that clearly, the tool isn’t ready for deployment.

When to move to Level 5: You find yourself always giving Codex the same feedback—“Don’t use this tone,” “Remember to add sources,” “Use the format from last time.”

Level 5: A System That Gets Better Over Time

At this level, your goal is that each piece of work makes the system better. Codex isn’t a one-off tool; it’s a system that accumulates experience.

After completing something valuable, make a habit of doing these four things:

1. Save effective prompts. Which prompt produced an especially good result? Save it exactly as is—what materials were used, what prompt was written, what format was requested, how it was checked. Reuse it directly next time.

2. Add mistakes to your checklist. Did numbers get wrong? Tone off? Made an unfounded assumption? Turn it into a checklist item, so Codex automatically checks against it in the future.

3. Update background files after project completion. After finishing a big project, update context.md—add new priorities, new tools used, what worked and what didn’t.

4. Ask Codex what’s worth consolidating. After each productive session, run this prompt:

Based on what we just did, what things could be turned into reusable templates, automated workflows, or small tools? What information should I add to the project files to make it faster to do similar things next time?

Part 4: 16 Workflow Templates

The following workflows are starting points for you. Adjust them based on your own tools and actual situation.

1. Inbox Zero

Suitable for: People who get anxious when email piles up, often miss replies they should have sent.

How to use: Have Codex scan your inbox for a certain period, categorize each email (needs reply / needs processing / can archive / already handled), draft replies for those needing response, clearly list actions needed for processing items, and separately flag things that require your own thinking. Only draft, don’t send—review and send in Gmail.

Get better over time: After a few runs, add a rules file—whose emails are always prioritized, which topics go straight to archive, what must be replied by you personally.

2. Daily Missed Messages Summary

Suitable for: People who switch between Feishu, email, and various groups, often forgetting who they need to reply to.

How to use: Have Codex find all Feishu messages sent to you in the past 24 hours that you haven’t replied to, and draft reply messages. Mark time-sensitive ones and those needing a thoughtful response. Automatic sending is not recommended; review and send manually.

3. Research Brief

Suitable for: People who need to quickly understand a topic before a meeting, planning session, or decision.

How to use: Have Codex compile a brief on a topic using links you provide, Feishu documents, Notion, Drive, and web search. Include: background knowledge, key data (with source links), differing viewpoints, questions you need to resolve, and three resources worth diving into. Mark uncertainties.

4. Write-and-Review – Writing Critique

Suitable for: Writing scenarios like WeChat public accounts, Xiaohongshu, project reports, etc.

How to use: Give the draft, style guide, and target audience to Codex. Let it first identify structural issues, factual errors, and tone problems, then make incremental changes. Don’t let it polish right away; have it diagnose first.

5. Writing Material Organization

Suitable for: People who need to sort out arguments, data, and counterpoints before starting to write.

How to use: Tell Codex your topic and core thesis, give it material links. Have it compile an “evidence bank”: strongest arguments supporting your point, strongest opposing arguments and how to address them, what evidence is still missing, and conflicting sources.

6. Text to Audio

Suitable for: People who prefer to learn by listening, or want to digest work content during their commute.

How to use: Have Codex convert text content (drafts, briefs, meeting minutes, reports) into an audio file and save it somewhere your phone can access (Dropbox, Drive, etc.). Listen while walking; jot down thoughts on your phone, process them later. Codex doesn’t have built-in TTS capabilities—you need to connect an external TTS API (e.g., Microsoft’s free TTS) or use a local TTS voice model.

7. Product Launch Plan

Suitable for: Pre-launch preparation.

How to use: Have Codex synthesize customer feedback, sales records, competitive materials, positioning documents, and historical launch experiences to generate a GTM plan. Output should include target users, positioning, messaging architecture, channels, timeline, risks, and items needing confirmation.

8. KPI Weekly Report

Suitable for: People who need to regularly pull data from multiple tools to generate reports.

How to use: Have Codex generate weekly or monthly KPI reports from analytics tools, revenue data, support data, and social metrics. Clearly define each metric’s definition, source, and time range. Flag abnormal fluctuations.

9. Customer Support Issue Triage

Suitable for: Teams that want to effectively convert customer feedback into product improvements.

How to use: Have Codex organize support tickets over a period, identify the core issue of each ticket, check if there’s already a record, link if yes, draft a new Issue if no. Prioritize frequently reported issues. Don’t create Issues directly; provide them for your review first.

10. Code Change Summary for Non-Programmers

Suitable for: Business stakeholders who raise development requirements.

How to use: Non-engineers can have Codex review a PR and summarize: the intent of the change, user impact, risks, and items that need product or operations confirmation. It doesn’t replace engineering review, but helps business stakeholders participate in reviews.

11. Candidate Shortlist

Suitable for: Recruitment scenarios.

How to use: Have Codex generate a shortlist of candidates based on the job description, resumes, interview records, and scoring criteria. Output should explain the rationale for recommendations, risks, gaps, and focus areas for the next interview round.

12. Strategic Planning / OKR

Suitable for: Managers who want to compress a few days of planning work into a few hours for quarterly planning or strategic review.

How to use: Have Codex compile a draft plan from past planning documents, meeting notes, and current data, listing assumptions, key choices, opposing views, and areas requiring human judgment. Require it to label what comes from source materials and what it added on its own.

13. Personal Learning Tool

Suitable for: People who want to use Codex to practice a skill.

How to use: Tell it what you want to learn, your current level, and how you want to practice. Have Codex build a small tool that works locally (e.g., chord identifier, language trainer). After each practice session, have it improve the tool based on your feedback.

14. Idea Collection Library

Suitable for: People doing content planning, product planning, or anyone who needs to continuously collect good ideas.

How to use: Have Codex maintain an idea library. For each idea, record its source, existing alternatives, value assessment, and next action. When adding a new idea, first check for duplicates against existing ones. Retain rejected ideas and the reasons for rejection.

15. Shared Dashboard Webpage

Suitable for: When reports, roadmaps, or review queues become too large to manage with documents.

How to use: When a project requires multiple people to continuously check status, comment, and review outputs, convert a document or dashboard into a Site. Sites are suitable for shared workspaces that need to track progress, not just for static reading. Site functionality is available only in Team and Enterprise editions.

16. Content Update Review

Suitable for: Cases where public pages, knowledge bases, or resource lists need regular updates and maintenance.

How to use: Have Codex compare existing content with the latest sources, list suggested changes, and provide reasons and evidence for each. Don’t automatically edit or publish. Place all suggestions in a review queue for your confirmation before publishing.

Part 5: Tips for Using Codex Well

How to “Manage” Codex

Using Codex is like managing a new hire on their first day who has great ability. You need to give them direction, resources, set boundaries, and check their output.

Tell it what result you want, not how to do it step by step. “Help me create a research brief using these sources, following this structure: first X, then Y” is much better than “First search this term in Feishu, then find that page in Notion, and then…”

For complex tasks, have it state its plan first. If a task will take several minutes or involve multiple tools, ask Codex to tell you its intended approach. This allows you to correct course before it goes off track.

Ask it what it still needs. Before starting, ask:

Before you begin, what else do you need to know to do a better job?

Require every important fact to be sourced. For documents that will be seen by others or used for decisions, every fact should be attributed. Add this rule to rules.md as a permanent rule.

Once the direction is confirmed, leave it alone. Only intervene if the direction is wrong, assumptions have changed, or it touches boundaries you haven’t approved. Otherwise, let it complete a full round before you review.

Always check results where they will actually be used. Something that looks okay in Codex might be completely off in Feishu, Slack, Gmail, or Notion.

Three Questions to Ask Before Checking Results

• What was the hardest decision you made while doing this?
• What alternatives did you consider but not use?
• Where are you least confident?

These three questions help you quickly find where Codex might have gone wrong.

Common Failure Scenarios

Confidently wrong. Codex might state incorrect facts with very confident language. Always verify all important factual content against original sources; don’t forward directly.

Numbers don’t match. Patching data from multiple systems can lead to inconsistencies—definitions differ, data is outdated, join methods are different. For numbers used in decision-making, verify each one individually.

Modified something it shouldn’t have. Codex sometimes “incidentally” changes files you didn’t ask it to. Especially for code-related tasks, review line by line what it actually changed, not just the final result.

Automation silently breaks. Tool APIs update, account passwords expire, your rules file becomes outdated—all can cause automation to fail. Every automation needs someone to check it regularly. “Set it and forget it” doesn’t work.

Misled by external content. Emails, web pages, and documents that Codex reads might contain instructions hidden from you (some not even visible to you). Therefore, put actions like “send, delete, modify” behind approval gates, and give Codex only the minimum necessary permissions.

How Teams Can Use It

As individuals get more comfortable, the team as a whole makes faster progress, but coordination is required.

A few key points: Each automation process must have a clear owner; Codex-generated documents should be placed in shared tools like Notion or Google Docs for everyone to review, not just viewed in Codex; write documents that are readable by both humans and AI, so everyone’s agents can use them.

The simplest way to promote it: Don’t roll it out to everyone at once. Find one person, one specific problem, demonstrate the results and process, and let others see, “This thing really saves time.”

Part 6: 7-Day Getting Started Plan

Day 1: Observe Your Work

List the tasks you repeat most often, find most annoying, or require the most information integration. Don’t think about automation yet; just identify all pain points.

Day 2: Connect Two Sources

Choose two critical systems (e.g., Feishu and a local Excel file) and have Codex complete a small task using them.

Day 3: Set Up Your Workspace

Create AGENTS.md, context.md, preferences.md, rules.md, and basic source files. Have Codex restate its understanding of your role, projects, and boundaries.

Day 4: Run a One-Time Task

Choose a low-risk but real task, such as organizing meeting notes or generating a research brief. Review Codex’s output and its uncertainties.

Day 5: Run a Multi-Source Task

Have Codex synthesize an output from multiple systems, like a weekly report, plan, or summary. Ask it to show you the plan first, then execute.

Day 6: Turn the Task into a Checklist

Write down the effective steps from previous days as a workflow document, including sources, output format, review criteria, and common mistakes.

Day 7: Decide Whether to Automate

If this workflow will recur, the risk is manageable, and the review criteria are clear, consider scheduling it. Otherwise, keep it as a manually triggered collaborative process.

Start Now

Codex’s value isn’t just “being able to write code”—it’s about becoming an intelligent workspace for knowledge work. The most important capability isn’t any single feature, but the cycle of connecting context, tools, processes, review, and reuse.

You don’t need to become an advanced user overnight. First, let it complete one real, low-risk, checkable task. Then save what worked. Then turn repetitive work into workflows. Finally, turn workflows into systems.

The real goal isn’t to have AI think for you—it’s to free up more of your attention for judgment, taste, strategy, and ultimate responsibility.

Open a folder in Codex. Pick one thing you would normally do anyway. Tell it what materials you have, what result you want, and how you’ll check. Do it well once, and it will naturally get smoother from there.