Big Idea

The argument in one line.

Developers can eliminate incompatible markdown schemas across AI coding tools by using a single .context folder synced to all platforms, paired with the PREVC workflow that requires human review before AI execution at every stage.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

A developer juggling 2+ AI coding tools (Claude, Cursor, Gemini, etc.) who manually recreates agent and context files in each tool's custom markdown format.
A software engineer who wants a single source of truth for AI agent configuration, documentation, and skills across multiple AI IDEs without manual syncing.
A developer familiar with CLI tools and MCP protocol who's willing to adopt a structured workflow (PREVC) to maintain human control over AI-generated code decisions.

SKIP IF…

You only use one AI coding tool — the fragmentation problem this solves doesn't apply to you.
You prefer graphical interfaces and are uncomfortable with terminal CLIs or configuring MCP connections in your AI agent.
You're looking for a tool that generates code end-to-end without human checkpoints — PREVC enforces review gates before AI writes anything.

TL;DR

The full version, fast.

AI coding tools each invented their own markdown schema for memory, agents, and skills, leaving developers juggling redundant files across Claude Code, Cursor, Codex, and Gemini. The fix is a single .context folder, managed by an open-source CLI and MCP server, that syncs symlinked agents, docs, plans, and skills into whichever tool you launch, so switching mid-task preserves state. Paired with the PREVC workflow (plan, review, execute, validate, confirm), the system enforces human checkpoints before code is written, replacing the typical "add authentication / 500 lines appear" loop with explicit approval gates. Adopt one folder as the source of truth, prompt the agent to plan before it codes, and you stay the operator instead of AI's copilot.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:31

01 · The nightmare

Pain named immediately: every AI tool has its own markdown memory schema, creating management chaos for developers using multiple tools.

00:31 – 01:05

02 · Solution intro

Introduces ai-coders/context, an open-source project with a single .context folder that syncs to all tools. MIT license. First English video disclaimer.

01:05 – 02:03

03 · Folder structure + sync

Demo of .context/agents, .context/docs, .context/plans structure. Quick sync command creates symlinks in Claude Code, Cursor, Codex folders. One change propagates everywhere.

02:03 – 02:51

04 · CLI + MCP setup

Two install paths: npx @ai-coders/context interactive wizard, or connect the MCP. Both analyze the codebase and generate context-aware agents, docs, plans.

02:51 – 03:43

05 · Plan workflow

After init, prompt plan [task]. The tool selects which agents to run, which docs to consider, and creates a cross-tool-portable execution workflow with persisted state.

03:43 – 05:52

06 · PREVC methodology

Five-phase structured AI workflow: Plan, Review, Execute, Validate, Confirm. Contrasted against autopilot AI that dumps 500 lines of unwanted code.

05:52 – 07:02

07 · Live demo

Real agent run in Portuguese. Three bugs discovered: TypeScript errors, tech debt, multi-tenancy risk. Three separate PREVC plans generated and linked to the workflow.

07:02 – 07:41

08 · CTA + close

Contribution ask: find bugs, write PRs. First English video acknowledgment. MIT license. AICoders channel intro.

Atomic Insights

Lines worth screenshotting.

The ai-coders/context CLI creates a single .context folder with a universal schema that syncs agents, docs, plans, and skills across Claude Code, Cursor, Codex, and Gemini simultaneously.
Each AI tool having its own memory file format (Claude.md, .cursorrules, etc.) creates a maintenance nightmare where the same information must be kept in sync across incompatible schemas.
A quick sync command maps the universal .context folder to each tool's native format via symlinks — changing one file updates all tools automatically.
The PREVC methodology enforces five steps: Plan, Review (iterate with AI until the plan looks correct), Execute, Validate (AI runs tests), Confirm (human deploys) — keeping the human as architect, not copilot.
The key failure mode of current AI coding workflows is moving from 'add authentication' directly to 500 lines of generated code — PREVC inserts a review step before any execution.
The plan phase surfaces dependencies, risks, and architecture decisions (OAuth vs JWT vs session, which providers) before a single line is written — the review phase validates them with the human.
A workflow memory file tracks exactly which step the build is on so you can switch from Claude Code to Codex to Cursor mid-task and continue from the same checkpoint.
The MCP version of ai-coders/context requires three steps: install the MCP, prompt the agent with 'init context', and the CLI generates agents, docs, and skills already mapped to your codebase.
The plan command generates not just tasks but a workflow of execution with specific agents assigned to each step — the plan is a sequence, not a list.
Tool-switching (from Claude Code to Codex to Cursor) without losing progress is the core value proposition — the .context folder is the portable state that makes this possible.
Three concurrent problems (type errors, stack depth, multi-tenancy risk) require three separate plans linked to the same workflow — parallel problem scoping without context collision.
The creator's observation that AI tools refuse to standardize memory file formats (because it would reduce lock-in) is the exact market gap that open-source context management tools fill.

Takeaway

Steal PREVC. Own your context.

Builder playbook

The developer who controls the spec controls the build -- PREVC is just that discipline made executable by AI.

PREVC is a named framework you can drop into any tutorial or pitch: Plan, Review, Execute, Validate, Confirm. Five phases, no ambiguity.
The .context folder model is directly applicable to JoeFlow CLAUDE.md -- one source of truth, synced everywhere via symlinks.
The add-authentication-then-500-unwanted-lines example is a universal pain point. Use it as a hook for any AI workflow content.
The you-are-the-copilot line is ready to clip at 4:31. Standalone, punchy, no setup needed.
Vini built for himself first, then opened it -- framing the CTA as help-fix-this-problem rather than subscribe. Worth mirroring for OSS and tool launches.

Glossary

Terms worth knowing.

CLI: Command-line interface — a text-based way to control software by typing commands into a terminal instead of clicking through a graphical interface.
MCP: Model Context Protocol — an open standard that lets AI assistants connect to external tools, data sources, and services through a shared interface.
Markdown file: A plain-text file using simple formatting marks for headings, lists, and links. AI coding tools commonly use them to store instructions, memories, and agent definitions.
Claude Code: Anthropic's command-line coding assistant that runs in the terminal and reads project-level instruction files to help write and edit code.
Cursor: An AI-powered code editor built as a fork of VS Code, with built-in chat, autocomplete, and agent features for writing software.
Codex: OpenAI's coding agent that runs tasks in the terminal or cloud, executing multi-step development work on a codebase.
Gemini: Google's family of AI models, offered as a coding assistant via a CLI and IDE integrations that compete with Claude Code and Cursor.
Antigravity: Google's agentic development platform that lets AI agents plan and execute coding tasks across an editor, terminal, and browser.
Agents: Specialized AI configurations with their own instructions, tools, and scope, invoked to handle specific tasks like code review, testing, or refactoring.
Skills: Reusable packages of instructions and scripts that extend an AI assistant's capabilities for specific tasks, loaded on demand when relevant.
Memory files: Markdown documents an AI tool reads at the start of each session to remember project context, preferences, and prior decisions across conversations.
Symlink: A filesystem shortcut that points one file path to another, so editing the original updates every linked location automatically.
Open source: Software whose source code is published publicly so anyone can read, modify, and redistribute it, usually under a permissive license.
MIT license: A permissive open-source license that lets anyone use, modify, and sell the software with almost no restrictions, as long as the original copyright notice is preserved.
Pull request: A proposed code change submitted to a shared repository for review and discussion before being merged into the main codebase.
OAuth: An authentication standard that lets users grant an app access to their account on another service without sharing their password.
JWT: JSON Web Token — a compact, signed token used to verify a user's identity between a client and a server after login.
Session auth: An authentication approach where the server creates a session record after login and tracks the user via a cookie on each request.
Multi-tenancy: A software architecture where one application serves many separate customers from a shared system, requiring strict isolation of each customer's data.
PREVC: A five-step AI coding workflow — Plan, Review, Execute, Validate, Confirm — that inserts human checkpoints before and after AI writes code so the developer stays in control.

Resources

Things they pointed at.

00:00toolClaude Code ↗

00:07toolCodex

00:07toolCursor AI ↗

00:40productai-coders/context ↗

05:52toolAntigravity (agent runner)

Quotables

Lines you could clip.

00:25

“Welcome to the nightmare of markdown files.”

Perfect title-echo punchline, standalone as a hook→ TikTok hook↗ Tweet quote

01:34

“One place, one folder to rule them out.”

Clean positioning line, LOTR reference lands naturally→ IG reel cold open↗ Tweet quote

04:31

“It is you that should be using AI as your copilot -- not you being the copilot of AI.”

The thesis of the whole video in one sentence, quotable standalone→ Newsletter pull-quote↗ Tweet quote

04:48

“User: Add authentication. AI: generates 500 lines of code. That is not what I wanted.”

Universal developer pain in three lines, comedic timing works→ TikTok hook↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphoranalogy

If you are a software developer like me, you have the same problem. You start using Cloud Code, and then you start using Codecs, and then you start using Gemini, Antigravity, Cursor AI, and all of those AI tools that we have right now.

The problem is each of them decided to implement their own way to make memories files, to create and manage agents, to deal with the skills.

So you are in the nightmare of markdown files. Welcome to the nightmare of markdown files. And I don't want you to live in this nightmare, so I have created the AI coders contacts to fix this stuff that all of these tools should be fixing right now.

But as I'm seeing, they don't want to do that. So let's do it together. In this video, you're going to see what I'm trying to do to fix that stuff for me and for you as well.

I have created an open source project to fix that nightmare. English is not my mother language. Disclaimer set.

Let's sign me to it. The proposal here is to have only the dot contacts folder. Inside of that dot contacts folder, you have your agents, your docs, your plans, your skills, your anything else they want to put a name in a markdown file organization.

After we have it configured here, it is pretty straightforward. You would just quick sync with the tool you want to use.

So if you are using CloudCode, quick sync with CloudCode, all your agents will have a sync link in your CloudCode folder.

So, basically, if you change the files inside of context folder, then it will automatically be recognized by cloud that this file has changed as well. Same for the other two. So you got the point.

One place, one folder to rule them out. I just want a comment that I can write in my terminal or an MCP that I can connect in my AI agent and say, hello.

I just want to initialize my context. And that's what I did. So if you prefer to use the CLI, there's an interactive CLI.

We have the AI agents and everything else that you need to have there. It will check your code base, analyze it, and then we'll suggest the next steps for you. Pretty straightforward.

You just need to hit enter enter enter, accept the terms that you never read, and there we go. Or you can just connect the MCP. With the MCP, there are some steps.

First one, install the MCP. Second one, prompt to the agent. Emit the contacts.

That's pretty much it. It will generate the files you just saw in my screen. But the thing is, these files, they are already mapped to understand your code base.

Your agents are for your code base. They are not, like, just general agents. They are not general skills.

They are not general docs. Next step, this will set up the context and field according with the code base. With the context ready, you need to prompt plan and then the task you want to implement.

So, for example, plan develop a new authentication method for me using AI context. This will create a new plan for you, but not just a random plan generated by AI.

It will line up the agents that need to be used for this. It will figure out what files of your documentation need to be considered for this plan, and, also, it will create a workflow of execution.

So if you start doing that in Cloud Code, you can continue doing that in Codecs. You can continue doing that in Corsair. No matter which tool you're using, if you want just to switch then, it's straightforward.

We have a memory that's persisted with exactly this step. It stopped it and which agent should go ahead and develop the feature in this step. Once you have planned the workflow, you have reviewed everything, that's pretty much it.

You can just hit start workflow and let AI work for you. And for the workflow system, I'm using something that I call Prevs. In Portuguese, it's Prevs.

It is a method that I have been using for years already since 2023 to develop software to use AI at all. Because first step, b, you plan.

R, you review your plan. So, basically, you will iterate with AI until your plan looks correct for you, until it looks legit, and then you move forward. Moving forward, we have three steps, execute, validate, and confirm.

So in the execution step, it will execute the code. In the validation step, it will validate for you. The idea here, it is not for you to run everything on autopilot.

The idea is for you to understand what's going on, to review step by step, to be assisted by AI. It is you that should be using AI as your copilot and and not you being the copilot of AI. Here's a quick example.

Most of AI coding workflows look like this. User, add authentication. Then AI generates 500 lines of code.

That's not what I wanted. That's you, user, and that's your workflow right now. But with this method, you will have p, what type of auth, OAuth, JWT, session, what providers?

R, here's the architecture. Dependencies, x, y, z. Risk is z.

Do you want to approve that? Do you want to move forward? You understand what's going on.

You understand the plan. You will review that with AI. Execution, implementing the approved design, implemented the approved feature.

Validation, all 15 tests are passing, and the security of it is passing, is complete. And see, it is deployed, or depending on your workflow, you can confirm and you can manually do that.

The idea is if you start doing something on Cloud Code and you are out of credits, you can just move away to Codecs and continue it exactly from where you were. Same thing. If you want to move to another tool, that's the same idea.

We'll have a memory file that will keep the track of the progress in your workflow. And here's an example on antigravity.

Basically, I just said, this is in Portuguese, in it is context. And then it started working, writing the files for me.

Next step, I wrote, write a plan to identify bugs and initiate an workflow in Previz. And then it started it and run every tools that were that were necessary, and then we have the plans.

As I was validating and checking the plans, I noticed that we had more than one problem. We had three problems.

The first one, we have type errors. The second one, attack depth.

And the third one, a most tenancy risk. Three different problems. Three different plans.

So I asked, hey. Let's generate three different plans for it and link it to the current workflow. And that's it.

That happened. And then I said, okay. Let's jump into the execution process and execute the plan.

And that's what Integraft is doing for me right now. I won't spend more time here. I want this to be as simple as the tool is.

So next step here is for you to help me to validate that tool, to find bugs, to write pull requests, to help us to fix this problem so our lives become better.

Okay? This is my first YouTube video in English. This is my first open source project that is actually becoming bigger than I imagined.

As I mentioned in the very beginning, when I started this, it was something it was something more for me and my colleagues, and then it started getting bigger. And now I'm organizing this in a structured way for collaborating with you guys. You're welcome.

If you have experience, you're welcome. If you want to just like four kids, you're welcome as well. This is MIT license, and this is AICoders channel.

I'm Vinny from Brazil. It's nice to meet you.

The Hook

The bait, then the rug-pull.

Every AI coding tool wants to own your memory. Claude Code gets CLAUDE.md. Cursor gets .cursorrules. Codex gets its own thing. Vini from Brazil got tired of maintaining them all separately, so he built a single source of truth and a five-step methodology to put humans back in control.

Frameworks

Named ideas worth stealing.

03:43acronym

PREVC Workflow System

P -- Planning
R -- Review
E -- Execution
V -- Validation
C -- Confirmation

A 5-phase structured AI coding methodology. No code is written until P and R are complete and human-approved. Forces spec-before-code discipline. Portable via the .context memory file.

Steal forAny pitch about structured AI workflows, anti-autopilot positioning, or the argument that developers should own their process

01:05model

ai-coders/context folder model

.context/agents
.context/docs
.context/plans
.context/skills

Single source of truth for all AI tool context. Sync command creates symlinks in each tool expected location. Change once, propagate everywhere.

Steal forCLAUDE.md architecture, multi-tool context management, OSS product positioning

CTA Breakdown

How they asked for the click.

VERBAL ASK

07:02product

“Next step here is for you to help me to validate that tool, to find bugs, to write pull requests, to help us to fix this problem so our lives become better.”

Soft community ask -- no subscribe push, no link card. Contribution-first CTA works well for OSS audiences.

MENTIONED ON CAMERA

00:00toolClaude Code ↗

00:07toolCursor AI ↗

00:40productai-coders/context ↗

FROM THE DESCRIPTION

OTHER LINKSAlso linked in the description.