17:02Andy Diep · Tutorial
ChatGPT + DaVinci Resolve: The Full AI Editing Setup
A video editor wires ChatGPT's Codex into DaVinci Resolve through an MCP server, then lets it build a full first-pass cut on its own, mistakes included.
August 7thThe argument in one line.
ChatGPT's desktop app is now three separate surfaces built for different jobs, and turning Codex into a real daily operating system means climbing a seven-step ladder from a scoped local folder to fully scheduled, self-directed work.
Read if. Skip if.
READ IF YOU ARE…
- You use ChatGPT mainly through one-off chats and want a repeatable system instead of re-explaining context every time.
- You're a solo builder, coach, or knowledge worker whose work already lives in folders of files, PDFs, scripts, or spreadsheets and want an agent to work on them directly.
- You've started automating AI tasks but keep getting inconsistent results because there's no durable project file or reusable skill behind them.
SKIP IF…
- You're looking for a review of raw model quality or benchmark comparisons, this is entirely about workflow structure, not model capability.
- You want a walkthrough of a different agent tool (Claude Code, Cursor, etc.) — the concepts transfer, but every example here is the ChatGPT desktop app specifically.
The full version, fast.
ChatGPT's desktop app has quietly become three separate surfaces: Chat for quick thinking, Work for finished deliverables, and Codex for work that touches real files on your computer. The video teaches a seven-rung ladder for using Codex as a daily operating system instead of a chatbot: start with a narrowly-scoped local project folder, encode only durable project rules in AGENTS.md, give each chat one clear job, turn anything you've explained twice into a reusable skill, add plugins only where they remove real friction, automate only work that's already stable by hand, and reserve Plan Mode and Goal Mode for larger multi-step jobs with a checkable definition of done.
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 →Where the time goes.
00:00 – 03:01
01 · The ChatGPT Desktop Update Explained
Aishwarya opens with why Codex has become her daily agentic harness, previews the seven-rung roadmap, and explains that Chat, Work and Codex are now one merged desktop app instead of two separate products.
03:01 – 05:17
02 · Chat vs. Work vs. Codex (Which to Use?)
The deciding question is the task type, not model choice: Chat for quick thinking, Work for finished deliverables built with sub-agents, Codex for work against real local files with full visibility into every change.
05:17 – 08:50
03 · Rung 1: Start with a Local Project Folder
Connect ChatGPT to the narrowest folder that holds the actual project context rather than uploading files repeatedly; anything Codex creates saves back into that same folder as a real, usable file.
08:50 – 12:34
04 · Rung 2: Teaching Your Agent with AGENTS.md
AGENTS.md should hold only durable, cross-session rules (tone, terminology, boundaries), built up from real corrections rather than written speculatively; nested AGENTS.md files in subfolders can override the root file.
12:34 – 14:23
05 · Rung 3: The "One Chat, One Job" Rule
Each chat should own a single outcome (research, script, fact-check, thumbnail) inside the same project and be named after its outcome; parallel agents editing the same file at once creates chaos, not speed.
14:23 – 19:04
06 · Rung 4: Turn Recurring Work into a Skill
A skill differs from AGENTS.md by scope: it teaches how to perform one specific recurring workflow, built only after doing the task successfully by hand once, then tested for correct triggering and boundaries.
19:04 – 22:58
07 · Rung 5: Safely Installing Plugins & Connectors
A connector grants authenticated access to one external tool, a plugin bundles connectors, skills and scripts together; every new plugin should start on a read-only task before being trusted with real actions.
22:58 – 25:09
08 · Rung 6: Automating Work with Scheduled Tasks
The safe build order is manual, then skill, then schedule; a standalone scheduled task spins up a new chat per run while an in-chat scheduled task reuses existing context, and both need narrow permissions.
25:09 – 27:46
09 · Rung 7: Using Plan Mode & Goal Mode
Plan Mode inspects files and asks questions before changing anything; Goal Mode runs a longer loop toward a stated outcome, constraints and a checkable definition of done, via an explicit produce-inspect-diagnose-fix cycle.
27:46 – 28:17
10 · Using Codex Remote on Your Phone
Codex Remote pairs the desktop session to a phone so actions can be approved and reviewed on the go, while the actual work still runs against the host laptop's own files and permissions.
28:17 – 29:21
11 · Summary & Free Codex Workbook
A recap of the seven-rung order plus a plug for the free downloadable Codex workbook that walks through the same sequence step by step.
29:21 – 31:09
12 · Bonus: Create Your Own Custom Codex Pet!
A walkthrough of Codex Pets, an animated status indicator for running, blocked, or ready tasks, including how to build a fully custom pet from an uploaded character image.
31:09 – 32:11
13 · Mastering Agentic AI Certification
Closing pitch for her six-week Mastering Agentic AI Certification program, covering LLM foundations, context engineering, RAG, agentic design patterns, fine-tuning, evaluations and AI security.
Lines worth screenshotting.
- ChatGPT's desktop app now has three separate work surfaces, Chat, Work and Codex, and only Work and Codex can spin up sub-agents.
- Codex isn't a coding-only tool: any file-based work, scripts, spreadsheets, decks, becomes agent-editable the moment it lives in a real folder Codex can touch.
- A local project should point at the narrowest folder that contains the work, not an entire home directory, because a smaller scope is both faster and safer.
- AGENTS.md should only ever contain instructions that would still be true if you opened a brand-new chat in that project next week.
- The single most reliable signal that something belongs in AGENTS.md is that you've already had to explain it more than once.
- A skill differs from AGENTS.md in scope: AGENTS.md governs how a whole project works, a skill governs how one specific recurring task gets done.
- Building a skill before you've completed the task successfully by hand skips the judgment calls the skill actually needs to encode.
- Plugins and connectors are not the same thing: a connector is authenticated access to one external tool, a plugin bundles several connectors, skills and scripts together.
- Every new plugin should start on a read-only task, summarizing an inbox instead of sending from it, before it's trusted with real external actions.
- The safe automation order is manual first, skill second, schedule third — scheduling a workflow that's never been stabilized by hand just runs the same mistakes unattended.
- A workable goal for Goal Mode needs a stated outcome, explicit constraints, and a checkable definition of done, not just a wish like "make it amazing."
- Correcting a plan takes five minutes; discovering an unplanned agent built the wrong thing beautifully costs forty-five.
- Custom Codex Pets are stored locally on the machine that made them and don't sync to the ChatGPT web app.
- Codex Remote lets you approve and review agent actions from a phone, but the work still executes against your host laptop's files and permissions, not a separate cloud environment.
How to actually run an AI agent instead of chatting with it.
AGENT WORKFLOW
A durable AI workflow is built in layers, scoped folder, durable project rules, single-purpose chats, reusable skills, permissioned plugins, tested automation, and only then bigger self-directed goals, each layer depending on the one below it being solid.
01The ChatGPT Desktop Update Explained
- The three modes (Chat, Work, Codex) split by task type, not by model quality — pick the mode based on what kind of output you need, not which one seems smartest.
- Chat handles quick conversational thinking, Work delivers a finished, sub-agent-assisted output, and Codex works directly against a real folder on your computer.
02Chat vs. Work vs. Codex (Which to Use?)
- Chat mode never spins up sub-agents; Work mode does — for anything moderately complex, the mode you pick changes how much horsepower is actually applied.
- Codex is not code-only: reports, spreadsheets, infographics and course material inside Codex still count as agent-editable work because they're real files in a real folder, not just chat text.
- The deciding question isn't which model to use, it's what kind of task this is: quick thinking, a deliverable, or live work against local files.
03Rung 1: Start with a Local Project Folder
- Connect the agent to the narrowest folder that actually contains the project's context, not your whole home directory — it's both faster and safer.
- A local project reads the current version of your files every time, so you stop re-uploading the same frozen PDF into a chat every few days.
- Anything the agent creates gets saved back into that same folder as a normal file, not trapped inside a chat response you'll lose track of.
04Rung 2: Teaching Your Agent with AGENTS.md
- AGENTS.md holds only what should stay true across many future tasks in that project, tone, terminology, approved sources, boundaries, never today's one-off instruction.
- Use this litmus test: would this instruction still apply if you opened a brand-new chat in this project next week? If yes, it belongs in AGENTS.md.
- Build the file from real corrections, not upfront guesswork — promote a fix to AGENTS.md only once you've had to give it more than once.
05Rung 3: The "One Chat, One Job" Rule
- One chat should own one outcome, research, script, fact-check and thumbnail work each get a separate, clearly-named chat inside the same project.
- Parallel agents working on separate drafts is useful; parallel agents editing the same file at the same time is chaos, not speed.
06Rung 4: Turn Recurring Work into a Skill
- A skill and a project's AGENTS.md answer different questions: AGENTS.md says how this project works, a skill says how to perform one specific recurring task.
- Build the skill after you've already done the task successfully once by hand — it needs your rejected drafts and judgment calls, not just the final output.
- Explicitly tell the skill what NOT to carry over (facts and examples specific to one run), or it will drag last week's topic into next week's unrelated task.
07Rung 5: Safely Installing Plugins & Connectors
- A connector gives access to an external tool, a skill gives the method, and a plugin bundles both together into one installable package, they solve different problems.
- Start every new plugin on a read-only task (summarize the inbox, don't send) before granting it permission to take real external actions.
- Install a plugin only when it removes friction you actually have this week, a long list of vaguely-interesting integrations just adds unused permissions.
08Rung 6: Automating Work with Scheduled Tasks
- The build order is manual, then skill, then schedule, automating a workflow you haven't stabilized by hand just runs the same mistakes unattended.
- A standalone scheduled task spins up a fresh chat each run (good for a daily brief); one inside an existing chat keeps that thread's context (good for monitoring one ongoing thing).
- Review the first several scheduled runs closely, if the output starts drifting, fix the underlying skill or prompt rather than hand-correcting each run.
09Rung 7: Using Plan Mode & Goal Mode
- Plan mode inspects the current files and asks clarifying questions before changing anything, use it whenever you don't already trust the approach.
- A workable goal states the outcome, the constraints, and a checkable definition of done; "make it amazing" gives the agent nothing to verify against.
- The actual work is a loop: produce the output, inspect it against the requirements, diagnose what failed, fix the smallest piece, then check again.
10Using Codex Remote on Your Phone
- Codex Remote pairs the desktop session with your phone so you can approve actions and review output without being at the laptop.
- The work still runs on your host machine's files, tools and permissions, remote just gives you a window into it, not a separate environment.
12Bonus: Create Your Own Custom Codex Pet!
- A visual status indicator (running / needs input / ready / blocked) is genuinely useful once you're juggling more than one agent task at a time.
- Custom pets are stored locally on the machine that created them, they won't follow the account to another device or the web app.
Terms worth knowing.
- AGENTS.md
- A durable instructions file inside a project folder that tells the agent rules which stay true across many future sessions, not instructions for the task at hand.
- Skill
- A reusable workflow package (name, description, process, references, assets, scripts) that teaches ChatGPT how to perform one specific recurring type of task consistently.
- Connector
- Authenticated access ChatGPT is granted to one external service, such as Google Drive, Gmail, Slack, or GitHub, so it can read or act on data there.
- Plugin
- An installable bundle that packages skills, connectors, MCP tools, hooks and scheduled-task templates together into one capability, such as Browser, Computer Use, or Sites.
- Local Project
- A ChatGPT project connected directly to one or more existing folders on your computer, so the agent always works with the current files rather than uploaded copies.
- Plan Mode
- A mode where the agent inspects the current context, asks clarifying questions, and proposes a sequence of steps before changing or creating anything.
- Goal Mode
- A mode for longer multi-step work where the agent keeps a stated outcome, constraints, and definition of done attached while it iterates, and can be viewed, edited, paused or resumed.
- Loop engineering
- The underlying practice behind Goal Mode: produce an output, inspect it against requirements, diagnose what failed, fix the smallest issue, then re-verify until the checklist passes.
- Codex Pet
- A small animated on-screen character that reflects a chat's live status, running, needing input, ready, or blocked, so multiple active tasks can be tracked at a glance.
- Scheduled task
- A recurring or one-time job ChatGPT runs automatically; a standalone task opens a new chat each run, while a task inside an existing chat reuses that thread's context.
Things they pointed at.
Lines you could clip.
03:05
“If you are somebody who is still using ChatGPT on the website version, stop being a dinosaur and let me fix this for you.”
16:20
“If I have explained the same process to chat GPT twice, I should probably stop explaining it and create a skill for it.”
22:30
“Scheduling a vague workflow does not make it better. It lets the vague workflow run without you watching it and that's actually dangerous.”
26:00
“I would rather spend 5 minutes correcting the plan than waiting 45 minutes and discovering that ChatGPT built the wrong thing beautifully.”
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.
metaphor
One of my biggest productivity hacks over the last several months has been working with Codex as my agentic harness. It has genuinely changed how I get work done. And because I have been living inside this product for months, I have also watched it evolve in real time.
New features show up, old ones get upgraded, and the models keep getting better. But one thing that has not changed is how productive it makes me. So I wanted to sit down and make you a detailed codex masterclass.
This is not a feature dump video. This is not one of those videos where I show you 30 buttons and by the end, you still have absolutely no idea what to do with any of them. I rather want to show you how I actually use this app to get real work done.
And trust me on this, I can call myself a power user based on how much I spend time on the application and how deeply I've dug into it. So sit back and relax because there's going to be a lot coming your way. You can also grab something to drink in the meantime.
Okay, now one quick thing before we dive in because the naming has shifted a little. Until a few weeks ago, there were two applications. One was ChatGPT and the Codex was a separate app.
Now as of July, OpenAI has combined these experiences into the updated ChatGPT desktop app. So now inside the same application, you will see ChatGPT Chat, ChatGPT Work and Codex. So when I say Codex Masterclass, it's not just about the Codex part of it, it's about the entire application.
But I just love the name Codex, so we're going to call it Codex. The way I'm going to teach you to learn about the application is using this . Trust me on this, there's a lot of pieces inside of ChatGPT and Codex now.
And if I can explain them in one giant list, it is going to get very confusing. So we'll start with the real folder on your laptop. Then we will add agents .markdown.
Then we'll organize your chats. Then we'll create skills. Then we bring in external tools through plugins.
Then we automate the stable workflows. And then finally use plan mode and goal mode for bigger tasks. You do not need to be a developer for any of this.
I use Codex for code but also for scripts, research, spreadsheets, presentations, website, course material and other operational work. So if your work lives in files, this setup will be useful for you. I actually call this category of people knowledge workers.
So if you're a knowledge worker as well, which basically means that you own a laptop and you work with multiple files on your laptop, this one is going to help you. Also, if you're new to my channel, hi there, I'm Aishwarya Srinivasan. I spent over 10 years working in machine learning and AI, including roles at IBM, Google and Microsoft.
And most recently, I was leading AI developer relations at Fireworks AI. And now I'm building my own AI company. while pursuing my passion to teach people how to become AI builders.
And Codex is genuinely one of my favorite tools which helps me wear multiple hats at the same time. Okay, so let's dive in. Now, if you are somebody who is still using ChatGPT on the website version, and let me fix this for you.
First, if you do not have the desktop application yet, go to chatgpt .com slash download and install it on your laptop. Now, once you open the new application, you will see chat. work and codex.
And the question you should ask is not which model should I use? The first question to ask is what kind of task am I trying to do? See, chat is where I go when I want to have a conversation.
Maybe I just want to understand the paper, brainstorm a few hooks, compare two tools, rewrite a paragraph or just ask a question. I'm just looking for a very quick answer or an opinion or something that I can continue thinking about. So basically, quick conversations, which may or may not necessarily need a broader context.
And maybe not something that I'm going to continue talking about later. So, let's say one time wins. Now, work mode is where I go when I want to hand off ChatGPT a larger piece of work and get a proper deliverable back.
Whether it's a research brief, a deck, a spreadsheet, a document, a website or a project plan. I care about the final result and I want ChatGPT to use the relevant tools, relevant files, connected sources and sub -agents to produce it. This is also a nuance that you need to know that ChatGPT chat version does not spin up subagents, but the work one does.
So if you're doing something slightly complex, go into the work section because it can actually spin up subagents to do it faster. Okay, then there is Codex. Codex is what I use when the task needs to work directly against a local folder or a code base on my laptop.
I can see the files it reads, the files it changes, the command it runs and the difference it's producing. It gives me much more visibility into the actual implementation. Now, please do not make the mistake of thinking Codex is only for code.
I create scripts, reports, spreadsheets, infographics, websites, and even course material inside Codex all the time. The difference is that those things exist as actual files inside a folder on my laptop. And I want the agent to work directly with them.
So my rule is very simple. If I want to think with chat GPT, I use chat. If I want a substantial deliverable and mainly care about the outcome, I use work.
And if the task depends on live local files and I want to see exactly what is being changed, I use codecs. It is pretty much the same application with overlapping intelligence, just difference in the working surfaces. Now that we have that out of the way, let's get into the ladder and get into the first rung.
So for the first rung, we start with a project and a real folder. So open the projects view inside ChatGPT desktop app. You will see ChatGPT projects and local projects.
A ChatGPT project can keep related chats, uploaded files, project instructions and connected sources together. A local project connects ChatGPT to one or more folders that already exist on your computer. So for the workflow that I'm showing you, the local project is where this gets extremely useful.
Say I'm creating a YouTube video. I already have a folder which contains papers, PDFs, screenshots, transcripts, brand assets, an outline of the video, and three versions of the script that I'm not ready to delete yet. If you're using ChatGPT the traditional way and you want to ask it questions based on all of these files, you'll do the dumb thing of just uploading all of this into ChatGPT.
And if you want to do something else with it again, three days later, you would do the same thing again. That's not the right way to work with a knowledge base. That's what a local project fixes.
I can connect ChatGPT directly into that folder. That means it works with the files currently sitting inside that folder without me requiring to upload it anywhere. Now, if I update the outline tomorrow and replace a PDF, the project can work with the current version on my laptop.
As I said, I'm not repeatedly uploading frozen copies. And this goes both ways. When Codex creates a script, a chart, a spreadsheet or a thumbnail brief, it can save the asset back into the same folder.
I can open it normally on my laptop. It is not trapped inside a chat response anymore. This is why I think non -developers are underusing Codex.
Your work may not be a GitHub repository, but it still lives inside folders full of PDFs, CSVs, presentations, images, notes, and templates. And Codex can work on top of all of that. Now, do not over -engineer your project structure.
Mine is intentionally boring. I will usually have sources, notes, assets, drafts, outputs, and archives. Just name your folder the way you want.
But just try to keep it clean so the model knows which folder to go into for which tasks. Now inside projects, I create a new local project and choose the specific folder that I want ChatGPT to work with. I'm not connecting my entire laptop.
I'm intentionally not giving it access to my home directory. I choose the narrowest folder that contains the context for that particular project. Because that actually saves the model to process too many documents.
This is both practical and safer. If I'm working on a Gen Academy video, Codex does not need access to an unrelated client folder or my personal documents. Then I give it a request such as review the papers in sources and working outlines in the note, create the first draft of my YouTube scripts in the draft, preserve the citations from my research notes, flag any claim that is not supported by the sources and do not modify anything which is inside the sources.
That is a useful prompt because I have told it where to look, what to create, where to save it, what must it protect and how to check the result. So use a project when the work will continue over time, where several chats need the same context or when you expect multiple outputs. If you're just asking one question about a random PDF, just attach the PDF.
You don't need to create a new folder for every thought that enters your head. But if you are building a course, managing a client project, running a content workflow, or analyzing a recurring data set, or just working on a product, give that body of work a proper project. Now, there is an interesting thing that you need to know.
ChatGPT has its own files. Now that ChatGPT has access to the files, the next question is, how do you stop yourself from repeating the same instruction in every single chat? This is where agents on Markdown come in.
That is our rung 2, which is teach the project how you want it to work using agents .markdown. Agents .md is not the prompt for the task that you're doing today. It contains the things that should remain true across many tasks inside that project.
So if you're working inside a code base, it might explain how the repository is organized, how to install dependencies, how to run the tests, which coding conversations to follow, what should never be changed. and what needs to pass before a task is finished. On the other hand, if you're using Codex for content, education, or operations, agents .md is still extremely useful.
It contains your audience, your tone, terminologies, approved sources, project conventions, brand rules, review checklist, and other boundaries that you want to specify. Now, here is how I recommend that you build this file. Don't sit down and try to write the perfect agents .md before you have any work done.
You're going to end up with a giant document full of aspirational nonsense that the agent cannot act on. Start with the obvious rules, then let your real feedback improve that file. Say ChatGPT fills a script with unnecessary analogies.
I would just go respond on the chat saying that remove these, I do not need random employee or kitchen analogies. Explain the product through the interface, a practical use case and what the viewer should do. That fixes the current output.
But the feedback does not automatically go into my agents .md. The chat knows it but the future chat may not. So if you want the correction to become a part of how this project works, I would have to go say based on the feedback that I just gave you, propose the smallest useful update to agents .md that would prevent this from happening in the future scripts.
And then I can actually just click on the agent's markdown and see what it changed. Codex will read my existing agent's markdown, tell me where the rule belongs and propose a targeted edit. I'll just review it and then I'll let it update the file.
Now this is the part that I really want you to use. As you add more characteristics to how you want the app to behave, convert the recurring ones to durable guidance. If I keep saying make my hooks shorter, that may need to become a rule.
If I repeatedly correct a technical term, add that into the approved terminologies. And if I always want three hook options with a visual and a verbal hook, just write it down. Now, please do not dump every piece of feedback into agents .markdown.
Something like make this paragraph 20 seconds shorter belongs to the script. While keep opening sections tight and avoid repeating the promise may be a durable writing rule. Here is the litmus test for what belongs inside the agent markdown versus what doesn't.
You should ask yourself, should this instruction apply again when I open a completely new chat inside this particular project next week? That should give you the answer. If yes, it probably belongs inside the agent markdown.
If no, it doesn't. You can also layer these files. The root agents .md can contain broad project guidance.
A file inside of a subfolder can have more specific rules and the closer file actually takes more precedence. So keep the main file short and put the detailed rubrics or checklist in a separate reference and point it to them. Now my favorite maintenance habit is this.
If the agent makes the same mistake twice, I do not only fix the output, I ask what durable instruction was missing and I update the project's agent .md. That is exactly how Codex gets better with you instead of you repeating yourself forever. Now, rung 3 is give each chat an actual job.
Now that we have the folder and we have a durable project instruction, the next layer is the chat itself. I know it sounds ridiculously simple, but it does make a huge difference. Use one chat for one outcome.
Do not create one giant conversation called YouTube and use it to research the topic, write the script, debate the title and create the thumbnail and integrate a sponsor and plan the distribution and analyze the video after it goes live. That chat is going to become a mess. Plus, it's not going to be efficient.
What I would rather do is for one video, I may have a research chat, a script chat, a technical fact -check chat, and a thumbnail chat. They all belong to the same project.
So they still have access to the relevant files and project guidelines. But each chat has a very clear job. I also name the chat after the outcome.
Something like research AI coding agents is much more useful than just calling it research. So when I come back two weeks later, I should know exactly what happened in that thread. And it's just much easier to search back later.
The main chat still owns the final result. I absolutely love parallel work, but there is a right way to do it. Several agents researching different questions is useful, but several agents randomly changing the exact same file is chaos.
For creative work, I ask them to return separate drafts, separate findings or review notes. Then the main chat decides what to use. For code, I isolate the changes using separate work trays.
So the rule is one chat, one outcome and clear ownership of the output. And once you have done a workflow successfully inside a chat, pay attention. If you know that you're going to do it again, you should not rebuild the entire process from scratch.
This is where we start talking about skill. So the rank four is if you will do it again, turn it into a skill. This is probably the feature that has changed my own chat GPT workflow the most.
If you don't know what's a skill, a skill is nothing but a reusable workflow that gives chat GPT task specific instructions, references, assets and optional scripts. It basically teaches ChatGPT how to perform a particular type of work in a consistent way. Now, if you're confused on was it similar to what agents .markdown is, let me give you a clear distinction.
Agents .md that we spoke about earlier says something like this is how we work inside this project. A skill on the other hand would say something like this is how you perform this particular workflow. My agents .markdown may say something like do not publish an unsupported technical claim.
But my YouTube script skill explains the entire process for creating a strong long form AI video, which includes researching the topic, creating hook options, building the narrative, adding practical demonstrations, checking the claims, reviewing for retention and formatting the final output. Now here is what I personally do.
Whenever I'm using ChatGPT to do something for the first time and I know that I'm going to come back and do the same thing later this week or in the next few weeks, I already start thinking about the skill that I can create based on that task. Now, I do not create the skill before I understand the workflow. So, I first complete the task.
I give it the feedback, I add the judgment that ChatGPT missed and I get the output to a point where I would actually want to use it. Then, I ask ChatGPT to turn that successful process into a skill. For example, you just need to type something like this.
Go through this entire chat and identify the reusable workflows that produce the final approved script. Create a skill for long -form AI educational videos. Include when it should trigger, the input it needs, the processes it should follow, the output structure, my writing preference, the technical review, and when it should stop and ask me for more information.
Do not carry over facts, examples or product details that belong only to this video. That last line matters. Otherwise, you build a skill for one codex video and three weeks later it is trying to insert codex pets into a video about rag.
You don't want to do that. So, a skill is usually a folder with a skill .md file. It starts with a name and description followed by the workflow.
The description tells ChatGPT when the skill should activate. Please do not write something vague such as helps with content. A good skill description should look something like this.
Use the skill when the user asks for a long -form educational YouTube video script about a certain topic. Do not use for short social captions, carousel copy or emails. Now, inside the skill, I define the purpose, the input, the workflow, the output format, the quality checks, the boundaries and what to do when something is missing.
And the references section can hold examples or any sort of rubrics. The assets section can hold templates and the scripts can handle deterministic checks such as validating a schema or checking image dimensions. And this is what makes skills so powerful.
So how do you actually create the skill? The example that I showed you earlier is actually using something which is built in called a skill creator where you just ask ChatGPT to build one skill for you based on the successful workflow within the chat that you're using. Or there's another option called record and replay where you perform a task and then it converts that into a skill.
Or the third option is just write down the markdown file yourself, which is going to be harder. So don't choose that first. Now, whichever route you use, do not accept the first draft blindly.
The agent can observe what you did, but it may not understand why you rejected a particular hook, why that particular source was unacceptable or what you checked before actually approving the output. That judgment needs to go into the skill. Then test it a few times.
Try a request where it should automatically activate the skill. Maybe some requests where it should be staying out of the way and a case where an input is missing. For a thumbnail skill, it should activate when I provide a headshot, stay quiet when I'm only using title ideas and request the image when I forget to attach it.
Now, this is an important thing. Compare the workflow with and without the skill. Did it trigger it correctly?
Did it protect the files? Did it reduce any correction rounds? And if it fails, identify whether the problem is the trigger or is it the input or is it the workflow or the quality rubric or maybe the permission to tools.
And then fix the smallest relevant part and then test it again. See, this is my actual rule. If I have explained the same process to chat GPT twice, I should probably stop explaining it and create a skill for it.
And if that skill also needs an access to an external service, that gets me to the next rung which is rung 5, plugins. This is the point where you open work mode or codex and then go to plugins. Plugins are available in work mode and codex but they are not available in the standard chat mode.
This is where the terminology can get a little messy so let me make it a little practical for you. See, a skill gives ChatGPT the method for doing a particular task. A connector gives ChatGPT authenticated access to tools or data in an external service such as Google Drive, Gmail, Slack, GitHub, Notion, or Figma.
A plugin is actually an installable bundle. It packages skills, connectors, MCP tools, hooks, assets, or other scheduled task templates together. And plugins go far beyond connecting tools like Gmail or Google Drive.
Browser lets ChatGPT open and interact with websites inside its built -in browser. Computer use goes a step further and lets it visually operate approved desktop applications on your laptop. Visualize turns a question or data set into an interactive chart, diagram, map, calculator, or simulation that you can explore inside the chat.
Sites lets you create and host a complete website, dashboard, game, or web application with a shareable URL. And Template Creator helps you turn a format you use repeatedly such as a report, handout, presentation, or brief into a reusable template. You do not need all of these.
Choose the plugin based on the workflow or output you're trying to create. For example, the Google Drive plugin lets ChatGPT work with your drive, docs, sheets, and slides through an authorized connection. It can also include the skills that define how to edit these files safely.
So, the connector gives you access, the skill gives the workflow, and the plugin packages the capability so that you can install and use it. Now, when you install a plugin, it may ask you to sign in to a relevant service. you still control the account and workspace that you authorize.
Installing a plugin is not a magical permission slip for ChatGPT to do whatever it wants inside your company's system. That's not how it works. All of these permissions can actually be controlled.
So here is how I decide what to install. I look at my actual day. Where am I repeatedly moving information by hand?
Downloading it, maybe uploading it into ChatGPT and copying the results back. If my research lives in Google Drive, it should probably connect to Google Drive. If I begin every morning reading my Slack channels and turning those updates into a priority list, Slack is also going to be a useful plugin that I should install.
If I'm asking Codex to work on pull requests, review issues, or understand repository context, then GitHub also becomes useful. If I'm constantly asking Codex to generate infographics for me, I might as well connect it with Canva and Figma. So what I would recommend is do not install everything that looks vaguely interesting.
You will end up with broader permissions that you need and no clear workflow for you to use any of it. So I would say start with the friction in your own work. And when you install a new plugin, begin with a read -only use case.
Say for example, you use Gmail plugin to find and summarize a set of emails before you actually give it permission in order to send one. Similarly, you can ask Slack to summarize a channel and draft a response before you give it access in order to post it. This lets you check whether ChatGPT found the right information and applied the right judgment before it takes external actions.
Now, you can describe the result and let ChatGPT choose the relevant plugin or type at in the composer and select it yourself. So, my recommendation is also to check whether a good plugin already exists before you build everything from scratch. Let's say your workflow is unique to how you work, then create a skill.
If you want to distribute that workflow, share it across a team or bundle it with external tools, package it as a plugin. Now this brings me to rank 6 which is automate the boring part but only after it is reliable. I feel automation is where everybody wants to start.
But that's not the right thing to do. They'll just install the application and immediately want ChatGPT to handle their inbox every single morning or create content while they're asleep or monitor their company and update a spreadsheet and somehow run their entire life before breakfast. Please, I beg you, do not start there.
My order is actually manual first, skill second and schedule third. I would say run the workflow yourself with ChatGPT. Fix the prompt.
Add the missing judgment. Turn that part into a skill. If it is something that you would reuse.
Then once you trust the output and it is something that needs to be repeated, say every single day or every single week or every two weeks, then schedule it. Scheduling a vague workflow does not make it better. It lets the vague workflow run without you watching it and that's actually dangerous.
So what you need to do is inside the app, open scheduled to create and manage recurring tasks. You can also ask chat GPT directly. Tell it what you want to run, when you want to run it and where the results should appear.
There are two patterns. A standalone scheduled task creates a new chat for every run, which is useful for things like morning brief or weekly report. And a scheduled task inside an existing chat returns to that particular conversation and uses its context, which is useful for monitoring a deployment, pull request or ongoing issues.
Now, if a scheduled task is working against a local project inside a desktop application, your host computer needs to remain powered on and available. For Git repositories, scheduled tasks can actually run inside dedicated work trees so they do not interfere with unfinished changes in your main checkout. Now for regular folders, the tasks work directly inside that project repository so be very deliberate about what it can edit.
You do need to understand that scheduled tasks run unattended. So keep the permissions narrow, test the exact prompts manually before scheduling them. Review the first few runs.
If the result starts drifting, fix the workflow or the skill instead of correcting every single output one by one. Now we're getting into another fun part which is rung 7 of using plan mode and goal mode for work that cannot fit into one response. This final rung is actually how I handle larger tasks where ChatGPT needs to investigate, make a plan, create several things, check the result and just keep going for several iterations.
Now if I don't trust the approach, I start with a plan mode. All you need to do is just type slash plan or use the plan mode shortcut. For example, I would say something like plan a system that turns one weekly AI research topic into a YouTube video, a newsletter and five short videos.
Inspect the current folder structure and existing examples. Ask me for the missing brand or publishing constraints. Identify which steps should stay manual, which could become a skill and which could eventually be scheduled.
Do not create or modify any files yet. Now plan mode can inspect the context. ask questions and propose the sequences before changing anything.
I would rather spend 5 minutes correcting the plan than waiting 45 minutes and discovering that ChatGPT . Now, once I agree with the approach, I use backslash goal for long running executions. The backslash goal is actually part of an engineering practice called loop engineering.
If you don't know what's loop engineering, go check out my previous video where I talk deeply about loop engineering. Now, A useful goal tells ChatGPT the outcome, constraints and the definition of done.
For example, something like create the complete launch package and output, preserve all source files, use only the approved statistics in verified sources .md. The task is complete when the long form script, news draft, 5 short video scripts and publishing checklist exist. Every technical claim has to be checked and all files need to pass the project review checklist.
The goal mode keeps the target attached while the work continues. You can view, edit, pause, resume or clear it. But goal mode is not going to risk you a terrible definition of done.
If you just say something like make my presentation amazing, that gives the agent nothing useful to verify. This is also where I think about the verification loop. You create the output, you inspect it, you compare it with the requirements, then you diagnose what failed, then you fix the smallest issue, then you verify it again.
You only stop when the checklist passes. For a deck, you create the slide, you render them, you check for weak hierarchy or clipped text, you fix them, you render it again. Same thing for code, you make a change, you run the test, you review the diff, you fix the regression and rerun the checks.
This is how you get from ChatGPT gave me something to ChatGPT completed the work. And if you need to step away from your laptop, you can even use it as remote. Set it up from your ChatGPT desktop app or your host computer and pair it with your mobile phone.
You can then start or continue the chat, ask questions, approve actions and review outputs right from your phone. I do that all the time. So if you see me walking around just constantly looking at my phone, most likely I'm working with Codex Remote.
This is amazing because actual work is still happening within your files, the tools, permissions and the environment which is connected to your host which is your laptop. Now I know this was a lot so let's just wrap it up.
I don't want to bring this all together because I don't want you to finish the video and install 20 plugins. What I want you to do is choose one body of work you will use again next week. Create a local project, connect to the specific project that you want to work on, add a short agents .markdown.
Then create a separate chat for each outcome. Then turn one repeatable workflow into a skill. Install a plugin only when it removes a real friction.
And then once the manual workflow is stable, schedule one safe recurring task. For the next larger project, you can try to use the plan mode before implementation and goal mode while it runs. I know when I'm explaining it right now, it sounds simple.
But when you're going to log into your ChatGPT app, I know you're going to get confused again and hence I have put together a very very detailed worksheet for you which is going to be a list of activities that you can just follow through and do it in front of you. I'm going to link that in the description below. Now before we close, I have to show you one of my favorite features on the app which is Codex Pets.
A Codex Pet is a small animated character inside the ChatGPT desktop application that shows what your chats are doing. A task may be running, need your input, be ready or become blocked. And you can see this reaction on that Codex pet.
I find it surprisingly useful because I usually have several tasks running at once. And I'm always multitasking on multiple different tabs on my laptop. So instead of going back and forth on Codex, I can actually glance at the pet and see that something needs my approval or something has finished.
It makes the app just feel more alive, which I love. And it's genuinely functional. Now to activate a Codex pet for yourself, there are a bunch of built -in pets.
So all you need to do is just go to your profile menu, go to settings, select pets and choose the pet that you want. Or you can even type slash pet inside the desktop application to wake it up. But this is the fun part.
You can actually create your own pet, which is what I have done. So inside settings, if you go to pets and select create your own pet, the app installs the bundled hatch pet skill. Then it reloads the skill and opens a new chat.
Inside that, you can describe the character that you want. What I have done is using ChatGPT images, I've created a cartoonish version of me and I've uploaded that. And ChatGPT creates and validates the required animated asset.
Then when it is finished, you can go back to settings, refresh the pet section and you'll actually see your custom pet right there. Now, one technical note, the custom pets that are created inside your desktop applications are stored locally on that computer. They're not going to be syncing into your chat GPT on the web.
Now, as I said earlier, if you want to play around with everything that I've covered in this video, I've created a free codex workbook which is linked in the description below. I would say don't try to explore every single feature at once. Try to go in the sequence that I've mentioned in the workbook.
And if this video made you want to go deeper and become an AR builder, do check out the Gen Academy. My co -founder Arvind and I have been building the Gen Academy to train individuals and companies in practical AI engineering. Our flagship program is Mastering Agentic AI Certification, which is a six -week deeply technical and practical certification where we cover generative AI and LLM foundations, context engineering, RAG, agentic design patterns, fine -tuning, AI evaluations, and AI security.
You'll actually be building projects every single week. You can either use low -code, no -code tools or code -heavy tools. or a combination that fits your background.
The entire point is to help both technical and non -technical professionals understand how AI systems work and how they are built. And if you end up building something after this masterclass, please do post it on LinkedIn or X and please tag me. I'm a huge advocate for building in public and I would genuinely love to see what you create and genuinely want to amplify your work.
and if you want more deeply practical ai videos from somebody who actually works with these tools every day do subscribe to my channel and i'll see you in the next
The Hook
The bait, then the rug-pull.
Aishwarya Srinivasan has spent months living inside ChatGPT's Codex, and she opens by promising the opposite of a feature dump: not thirty buttons you'll forget, but a seven-rung ladder for actually running your work through it, starting with a folder on your laptop and ending with agents that plan, execute and check their own work.
Named ideas worth stealing.
04:10list
The Seven-Rung Ladder
- Rung 1: Start with a local project folder
- Rung 2: Teach the agent with AGENTS.md
- Rung 3: One chat, one job
- Rung 4: Turn recurring work into a skill
- Rung 5: Install plugins and connectors only where needed
- Rung 6: Automate with scheduled tasks, after it's stable
- Rung 7: Use Plan Mode and Goal Mode for larger work
A seven-step progression for turning ChatGPT's desktop app into a working daily system instead of a one-off chatbot, with each rung depending on the durability of the one below it.
Steal foran onboarding checklist for a team adopting an AI agent tool, or a personal audit of how far your own agent usage has actually progressed
How they asked for the click.
VERBAL ASK
31:09product
“Our flagship program is Mastering Agentic AI Certification, which is a six-week deeply technical and practical certification where we cover generative AI and LLM foundations, context engineering, RAG, agentic design patterns, fine-tuning, AI evaluations, and AI security.”
soft pitch placed at the very end, after the free workbook has already been given away, framed as a next step for going deeper rather than a hard sell
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
