@blueskylh1: The most painful thing about solo product development or leading an AI team is being a "mindless messenger" between different chat windows. After the PM writes the requirements, I have to copy and paste them into the developer's chat. After seeing the sharing from Jason @jxnlco, a developer experience engineer on the OpenAI Codex team, I set up a workflow without...

X AI KOLs Timeline Tools
ai-workflow collaboration developer-experience agent-team codex

Summary

Introduces a multi-AI agent collaborative workflow based on local plain text files and OpenAI Codex, allowing PM, backend, frontend, and QA to efficiently develop via file relay without copy-pasting.

The most painful thing about solo product development or leading an AI team is being a "mindless messenger" between different chat windows. After the PM writes the requirements, I have to copy and paste them into the developer's chat. After seeing the sharing from Jason @jxnlco, a developer experience engineer on the OpenAI Codex team, I got a workflow that allows PM, backend, frontend, QA, and AI to collaborate efficiently using only local plain text files, without any complex code. This approach also works for different agent teams. (Using the development of an "Automated Overseas Tech News Weekly Tool" as an example, here is a complete step-by-step practical tutorial.) **Step 1: Set up a local file cabinet** Create a main folder `vault/` on your local computer or cloud drive, and create four empty subfolders inside: - `vault/projects/` (for core technical handover documents) - `vault/people/` (for team or contact requirements) - `vault/notes/` (for temporary memos and drafts) - `vault/agents/` (for AI employee handbooks) **Step 2: Add the master rule (AGENTS.md) to the root directory** Create a new file `AGENTS.md` in the root of `vault/` and directly copy and paste the rules from the comment section below. This is for all AI. (The specific content is posted in the comments section.) **Step 3: Distribute four job descriptions in agents/** In the `vault/agents/` directory, create four corresponding `.md` files and paste the following responsibilities: - **pm.md (Product Manager):** Responsibilities: Turn the owner's ideas into requirements. Has write access to `projects/`. Create and update `projects/product_spec.md` (requirements document), define features and acceptance criteria. - **backend.md (Backend Expert):** Responsibilities: Write core logic and APIs. Read-only access to requirements documents. Create and write `projects/api_design.md` (API document, define input/output parameters), and write core crawling and summarization code. - **frontend.md (Full-stack Frontend):** Responsibilities: Handle UI design and frontend code. Read both the requirements document and backend API document. You need to design the visual appearance of the entire UI, Tailwind layout, and write the interface code. Strictly avoid inventing data fields that don’t match the backend. - **qa.md (Test Expert):** Responsibilities: Run tests, catch bugs. Read all code and documents from the full pipeline. Create and write `projects/test_plan.md`, report unfixed bugs and tag frontend/backend for fixes. **Step 4: Run the development pipeline** Mount the `vault` folder in your AI tool (e.g., Codex), open 4 pinned conversations (Thread 1 to 4), and start the relay: 1. **Wake up the PM to gather requirements (Switch to Thread 1)** Instruction: "Read AGENTS.md and agents/pm.md. You are now officially the PM. I want to build an automated overseas tech news weekly tool. The core is to fetch hot posts on X and summarize them in plain geek language. Go." AI action: It will think in the background, automatically generate `projects/product_spec.md` locally. 2. **Wake up the backend to write logic (Switch to Thread 2)** Instruction: "Read AGENTS.md and agents/backend.md. Start as backend. Go read the PM's file, define the backend API for this weekly tool, and write the core crawling script." AI action: It reads the PM's document, writes `projects/api_design.md` locally, and finishes the Python crawling logic. 3. **Wake up the frontend to build the UI (Switch to Thread 3)** Instruction: "Read AGENTS.md and agents/frontend.md. Start as frontend. Go read the requirements and backend API, design the UI visual style and tech color scheme for the weekly page, then pixel-perfect implement with Next.js, matching the APIs." AI action: The frontend enters, holding requirements in one hand and backend API in the other. It designs the news card style and seamlessly stitches frontend and data with Tailwind CSS. 4. **Wake up the QA to catch bugs (Switch to Thread 4)** Instruction: "Read AGENTS.md and agents/qa.md. Start as QA. Go read the code and documents written by the previous team members, run automated tests, and report bugs in a test report." AI action: QA automatically runs tests. If long text overflows, it creates `projects/test_plan.md` locally with the record: [BUG #01] Text not truncated, causing overflow. Owner: @Frontend. Later, let the frontend take a look and fix it. Throughout the process, as the CTO, you only need to give orders in 4 windows without copying and pasting a single word of requirements or code. They collaborate asynchronously through local text building blocks. If this workflow helps you, feel free to follow, like, share, and comment so more agent teams can see it.
Original Article
View Cached Full Text

Cached at: 06/02/26, 07:36 PM

When building a product solo or leading an AI team, the most painful part is acting as a “mindless messenger” between different chat windows. The PM writes requirements, then you have to copy and paste them into the developer’s chat box.

After reading a share from Jason @jxnlco, a Developer Experience Engineer on the OpenAI Codex team, I got a workflow running that requires no complex code—just local plain text files—allowing PMs, backend devs, frontend devs, QAs, and AI to collaborate efficiently. This approach also works for different agent teams.

(Taking the development of an “Automated Overseas Tech News Weekly Tool” as an example, here’s a complete step-by-step practical tutorial.)

Step 1: Set Up Local File Cabinet

Create a main folder vault/ locally or on cloud storage, with four empty subfolders inside:

  • vault/projects/ – for core technical handover docs
  • vault/people/ – for team or contact requirements
  • vault/notes/ – for temporary memos and drafts
  • vault/agents/ – for AI employee handbooks

Step 2: Place the Master Guide (AGENTS.md) in Root

Create a file AGENTS.md in the vault/ root directory. Copy and paste the rules from this comment post directly into it — this is for all AIs.

(The specific content is placed in the comments section.)

Step 3: Distribute Four Job Descriptions in agents/

Under vault/agents/, create four corresponding .md files and paste the following responsibilities:

  • pm.md (Product Manager):
    Responsibilities: Turn the owner’s ideas into requirements. Has write access to projects/. Create and update projects/product_spec.md, define features and acceptance criteria.

  • backend.md (Backend Expert):
    Responsibilities: Write core logic and APIs. Read-only access to requirements docs. Create and write to projects/api_design.md (define input/output parameters), and write core scraping and summarization code.

  • frontend.md (Full-stack Frontend):
    Responsibilities: Handle both UI design and frontend code. Read both the requirements doc and backend API doc. Design the entire UI’s visual appearance, Tailwind layout, and write the interface code. Strictly forbidden to invent data fields that don’t match the backend.

  • qa.md (Test Expert):
    Responsibilities: Run tests, find bugs. Read all code and docs from the full pipeline. Create and write to projects/test_plan.md, list failing bugs and tag frontend/backend to fix.

Step 4: Run the Development Pipeline

Mount the vault folder into your AI tool (e.g., Codex), open 4 pinned chat threads (Thread 1 to 4), and start the relay:

  1. Wake PM to write requirements (Switch to Thread 1)
    Instruction: “Read AGENTS.md and agents/pm.md. You are now the PM. I want to build an automated overseas tech news weekly tool. Core: scrape popular posts from X and summarize in geek-friendly plain language. Go.”
    AI action: It thinks in the background and automatically generates projects/product_spec.md locally.

  2. Wake backend to write logic (Switch to Thread 2)
    Instruction: “Read AGENTS.md and agents/backend.md. Act as backend. Go read the PM’s file, define the backend APIs for this weekly tool, and write the core scraping script.”
    AI action: It reads the PM’s doc, writes projects/api_design.md locally, and finishes the Python scraping logic.

  3. Wake frontend to build UI (Switch to Thread 3)
    Instruction: “Read AGENTS.md and agents/frontend.md. Act as frontend. Go read requirements and backend APIs. Design the weekly page’s UI visual style and techy color scheme yourself, then pixel-perfectly implement it with Next.js and wire up the APIs.”
    AI action: Frontend steps in, taking requirements in one hand and backend APIs in the other. It designs news card styles and seamlessly stitches the frontend with data using Tailwind CSS.

  4. Wake QA to catch bugs (Switch to Thread 4)
    Instruction: “Read AGENTS.md and agents/qa.md. Act as QA. Go read all the code and docs from the previous people, run automated tests, and report bugs in a test report.”
    AI action: QA runs tests automatically. Finds long text overflow — creates projects/test_plan.md locally with record: [BUG #01] Text not truncated causing overflow. Owner: @Frontend. You then ask frontend to take a look and fix it.

Throughout the entire process, as CTO, you only need to give orders in 4 windows — no need to copy and paste a single word of requirements or code. They all collaborate asynchronously through local text blocks.

If this workflow helps you, feel free to follow, like, share, and comment so more agent teams can see it.

Similar Articles

@xiaogaifun: A Major Pitfall Our Team Encountered with Codex. I want to sincerely share a topic we discussed in our weekly meeting. This might be a common pitfall for many people using agent products like Codex. In short, if we just keep issuing commands and accepting suggestions from Codex without understanding or judgment, then these...

X AI KOLs Timeline

The team shares the problem of over-reliance on AI leading to a lack of deep understanding among members when using Codex. It emphasizes that humans should act as leaders of AI, not mere messengers, and need to understand the principles and logic, turning AI output into their own experience.

@justloveabit: With This Open-Source Tool, I Got a Team of AIs to Work for Me. Here's the deal: I've been tinkering with various AI agents lately. Multiple Claude Code windows open, Codex running, occasionally using Cursor. The result? Total chaos—I had no idea what each agent was doing or how much it was costing. Restar…

X AI KOLs Timeline

This article introduces Paperclip, an open-source tool designed to centrally manage and orchestrate multiple AI agents. By simulating a corporate organizational structure, task assignment, and budget control, it addresses key pain points in multi-agent collaboration, such as lost context, unpredictable costs, and chaotic scheduling.

@dotey: https://x.com/dotey/status/2057250417638035555

X AI KOLs Timeline

This article shares usage tips from the Codex official team, including persistent conversation flow, voice input, task intervention and queuing, tool integration, automation, and goal setting, to help users get the most out of Codex, an AI coding agent.

@php_martin: https://x.com/php_martin/status/2064975977860440439

X AI KOLs Timeline

This article provides a comprehensive introduction to the features and usage of the OpenAI Codex desktop application, including project management, skill/plugin system, automation, and multi-task parallel development strategies. It also offers practical cases and risk warnings, aiming to help users efficiently utilize AI agents for parallel development.

@AdrianPunk115: The role of Codex in the entire workflow: Project manages context, Thread manages tasks, Plan manages the route, Goal manages execution, Skill manages reuse. When you start treating it as an AI workbench that can take over the complete workflow, rather than a chatbot, many functions suddenly come together…

X AI KOLs Timeline

Explains how Codex functions as an AI workbench that integrates project context, tasks, plans, goals, and skills into a cohesive workflow, positioning it as more than just a chatbot.