How to Do Everything Claude Does With ChatGPT: Skills, Agents, Subagents, and Automation

Every week I see another video claiming Claude can now replace an entire team.

Install this Skill.

Create five subagents.

Connect an MCP server.

Give each agent a different job.

Type one command.

Then supposedly your AI employees research competitors, find leads, write content, analyze data, build software, and return finished work.

Some of these demonstrations are genuinely impressive.

But after seeing enough of them, I started asking a different question:

Wait. Can’t ChatGPT already do most of this?

I use ChatGPT constantly, so I decided to look past the viral demos and compare the actual systems.

What I found surprised me.

Claude and ChatGPT increasingly share many of the same underlying ideas: reusable Skills, persistent project instructions, specialized subagents, parallel agents, external tool connections, MCP, plugins, scheduled workflows, and developer agent frameworks.

Claude frequently presents these ideas differently. In some areas, Claude makes the structure easier to see. In a few areas, Claude has capabilities that do not map perfectly to a ChatGPT button.

But if your real question is:

“Can I build the kinds of AI workflows people are showing with Claude using ChatGPT?”

For most of the interesting use cases, the answer is yes.

In this guide, we are not going to debate which chatbot is smarter.

We are going to build something.

By the end, you will know how to turn ChatGPT from a chatbot you ask questions into something closer to an AI marketing department with reusable procedures, specialist workers, connected tools, recurring jobs, and one director coordinating the whole operation.

---

Claude vs. ChatGPT: The Translation Guide

Before building anything, it helps to translate Claude terminology into OpenAI terminology.

Claude concept	ChatGPT/OpenAI counterpart
Claude Skills	ChatGPT/Codex Skills
CLAUDE.md	Project Instructions / AGENTS.md
Claude Projects	ChatGPT Projects
Claude subagents	ChatGPT Work / Codex subagents
Claude Agent Teams	Parallel subagents / multi-agent systems
Claude Connectors	ChatGPT Apps
Claude MCP	MCP in Codex/OpenAI agent systems
Claude Plugins	ChatGPT/Codex Plugins
Claude Code	Codex
Scheduled workflows	ChatGPT Scheduled Tasks / Codex Automations
Claude Agent SDK	OpenAI Agents SDK

This table needs one qualification:

These are conceptual counterparts, not always identical products.

For example, Claude Agent Teams are more explicit than ordinary ChatGPT subagent workflows. Claude Hooks have no perfect one-click equivalent in normal ChatGPT.

Yet the overall architecture has become remarkably similar.

Anthropic defines Claude Skills as folders containing instructions, scripts, and resources that Claude loads for repeatable specialized tasks. OpenAI now describes ChatGPT and Codex Skills in much the same way: reusable workflows containing instructions, resources, and optional scripts. (Claude Help Center)

That brings us to our first build.

---

Step 1 — Create a ChatGPT Project That Acts Like Your AI Headquarters

If you use ChatGPT for serious ongoing work, stop starting everything from unrelated blank chats.

Create a Project.

ChatGPT Projects can group chats, reference files, instructions, memory, and related work under one objective. OpenAI describes Projects as smart workspaces for repeated or evolving work such as research, planning, and writing. (OpenAI Help Center)

For our example, create:

AI Marketing Department

This becomes the headquarters for everything we build in this article.

Inside the Project, you might upload or provide:

• descriptions of your products
• audience profiles
• customer questions
• past articles
• sales pages
• brand guidelines
• competitor notes
• keyword research
• successful emails
• offers
• testimonials
• internal procedures

You are trying to create a body of context that future work can draw from.

Add Project Instructions

Give the Project rules such as:

You are part of my AI marketing department.

Our goal is to identify worthwhile marketing opportunities and turn them into useful publishable assets.

Prioritize evidence over hype.

Separate verified facts from assumptions.

Write for entrepreneurs and marketers who want practical instructions.

Use strong curiosity-driven openings.

When researching software, verify current capabilities before making claims.

Give step-by-step instructions when the reader needs to perform an action.

Flag anything that requires human approval before publication.

That one setup changes the relationship.

You are no longer explaining your business from scratch every time.

TRY THIS IN CHATGPT

I want to turn this Project into an AI marketing headquarters. Help me identify the permanent information you should know about my business, audience, products, competitors, brand voice, content standards, and approval rules. Organize what I give you into a concise operating manual for this Project.

That is the easy, nontechnical starting point.

Now let’s move into something closer to Claude’s famous CLAUDE.md.

---

Step 2 — Give ChatGPT Permanent Operating Instructions

Claude Code users often use a file named CLAUDE.md.

Its purpose is simple: give Claude project-level instructions it should read when working.

Codex has its own version of this idea:

AGENTS.md

OpenAI states that Codex reads AGENTS.md files before doing work. Instructions can be layered, letting broader project rules apply across a workspace and narrower instructions apply closer to particular files or tasks. (OpenAI Developers)

This is an excellent concept whether you write code or not.

Your AGENTS.md might say:

# AI Marketing Operating Rules

## Audience
Write for entrepreneurs, marketers, creators, and small business owners.

## Research
Verify current software claims using reliable sources.
Separate facts from assumptions.
Do not invent statistics.

## Content
Begin with a curiosity-driven hook.
Use practical examples.
Give concrete instructions.
Avoid filler.

## SEO
Use keywords naturally.
Prioritize usefulness over repetition.
Include search intent in article planning.

## Quality Control
Check factual claims.
Check that instructions still match current interfaces.
Identify any claims that need human review.

## Final Output
Return work in publication-ready form.

Think about what has happened here.

Previously, you might have copied this giant instruction into every prompt.

Now it becomes infrastructure.

CLAUDE.md vs. AGENTS.md

The names differ.

The idea is similar:

Teach the AI how work should be performed before asking it to perform the work.

That distinction matters.

Most beginners spend all their time improving individual prompts.

Advanced AI workflows increasingly move instructions upstream into reusable systems.

---

Step 3 — Turn Repetitive Prompts Into Reusable ChatGPT Skills

This may be the biggest misconception created by the current Claude craze.

People see someone installing a Claude Skill and assume Claude has invented a capability ChatGPT lacks.

ChatGPT now has Skills too.

OpenAI describes a Skill as a reusable workflow that can package instructions, resources, examples, and optional scripts. Skills can be used by ChatGPT and Codex for task-specific work. (OpenAI Help Center)

So what should become a Skill?

Anything you repeatedly explain.

Suppose every time you research a competitor you say:

1. Identify their core offer.
2. Identify their target customer.
3. Find their main traffic sources.
4. Analyze messaging.
5. Find customer complaints.
6. Identify gaps.
7. Suggest opportunities.

Stop typing that repeatedly.

Turn it into a Skill.

Example: Competitor Intelligence Skill

A simplified SKILL.md might look like this:

# Competitor Intelligence

When asked to analyze a competitor:

1. Identify the company and primary offer.
2. Identify likely target customers.
3. Review positioning and messaging.
4. Identify major products and pricing when available.
5. Find recurring customer praise.
6. Find recurring complaints or unmet needs.
7. Identify gaps competitors appear to ignore.
8. Rank the three most interesting opportunities.
9. Separate confirmed evidence from inference.
10. Provide source references for current claims.

You could build Skills for:

• competitor research
• keyword opportunity analysis
• YouTube research
• sales-page reviews
• lead qualification
• customer review mining
• content briefs
• email campaigns
• financial modeling
• product research

OpenAI’s Skill documentation says Skills can package instructions, resources, and scripts, rather than being limited to one giant prompt. (OpenAI Developers)

That distinction is worth remembering:

A prompt tells AI what to do once.

A Skill teaches AI how you want a recurring job performed.

TRY THIS IN CHATGPT

I repeatedly perform the following job: [describe your workflow]. Turn this process into a reusable Skill. Identify the trigger, required inputs, step-by-step procedure, quality checks, output format, and any reference files that would make the Skill more reliable.

Once you start thinking this way, you will probably find ten workflows in your business that should stop living as loose prompts.

---

Step 4 — Build Your First ChatGPT Subagent Team

Now we reach one of the features driving the Claude excitement.

Subagents

A subagent is simply a specialist worker given a bounded assignment.

Claude Code supports custom subagents with their own instructions, tools, and context. (Claude Platform Docs)

OpenAI now supports subagent workflows in ChatGPT Work and Codex. OpenAI describes them as specialized agents that can work in parallel before their findings are collected into a main response. (OpenAI Developers)

This matters for one simple reason:

A giant prompt asking one AI to be researcher, strategist, critic, writer, and fact checker at once is often weaker than dividing the work.

Let’s create four roles.

Subagent #1 — Market Researcher

Its assignment:

Find customer problems, market shifts, new demand, changing behavior, and emerging opportunities.

Subagent #2 — Competitor Analyst

Its assignment:

Study existing solutions, positioning, pricing, promises, complaints, and gaps.

Subagent #3 — SEO Opportunity Researcher

Its assignment:

Identify searches that reveal commercial intent, unmet questions, and content opportunities.

Subagent #4 — Offer Strategist

Its assignment:

Examine the first three reports and identify offers people might actually pay for.

Then give the main agent the director role.

TRY THIS IN CHATGPT

Act as the director of an AI marketing research team. Delegate this opportunity analysis to separate specialist subagents when the available mode supports subagents:

1. Market Researcher — identify demand and customer problems.
2. Competitor Analyst — identify existing offers and gaps.
3. SEO Researcher — identify search behavior and unanswered commercial questions.
4. Offer Strategist — identify monetizable solutions.

Have the specialists work independently rather than copying one another’s assumptions.

Then synthesize their findings into one report. Show where their conclusions reinforce one another, where they conflict, and rank the top three opportunities.

Notice the instruction:

Work independently.

That is deliberate.

If every agent sees everyone else’s answer first, you can create artificial consensus.

Independent research followed by synthesis is often more interesting.

---

Step 5 — Turn Four Agents Into a Seven-Person AI Marketing Department

Now we extend the system.

Your AI department becomes:

1. Trend Researcher

Finds rising topics, changing customer behavior, new tools, and emerging problems.

2. Keyword Researcher

Studies what people search for and what those queries reveal about intent.

3. Competitor Analyst

Studies offers, messaging, pricing, strengths, weaknesses, and market gaps.

4. Article Writer

Turns the research into a useful article rather than merely restating research notes.

5. Fact Checker

Checks claims, dates, names, software capabilities, statistics, citations, and unsupported conclusions.

6. Visual Director

Creates image concepts, diagrams, featured-image ideas, screenshots needed, and visual storytelling opportunities.

7. Distribution Strategist

Turns the final asset into email, social, video, newsletter, and promotional opportunities.

Above them sits:

The Marketing Director

The director does not need to perform every job.

Its job is to decide:

• what needs to be done
• which specialist should do it
• what order makes sense
• what can happen in parallel
• which findings matter
• where conflicts need resolution
• when human approval is required

That is an agentic workflow.

It has almost nothing to do with pretending AI is a human employee.

It is simply structured delegation.

---

Step 6 — Connect ChatGPT to the Tools You Already Use

Agents become more useful once they can work with your real information.

ChatGPT Apps provide connections to external services and data. Since July 2026, OpenAI has organized workflow discovery through its Plugin Directory; plugins can include Skills, Apps, and app templates. Apps remain the part that connects ChatGPT or Codex with external data or actions. (OpenAI Help Center)

Depending on your available connections, this can include services used for:

• email
• calendars
• documents
• design
• contacts
• development
• business data
• financial information

This is where the meaning of “agent” changes.

Without tools, your agent can advise.

With tools, it may be able to retrieve information or take permitted actions.

MCP

MCP, or Model Context Protocol, is another connection layer that lets AI systems work with external tools and data sources.

Claude has heavily promoted MCP in its agent ecosystem.

Codex and OpenAI’s developer stack support MCP-style tool connections too.

For most business users, you do not need to become an MCP engineer before receiving value.

Start with supported Apps and Plugins.

Move into MCP when your workflow requires a system that does not already have an easy connection.

Read vs. Act

This is a distinction every user should learn.

A tool may let an agent:

Read

• search documents
• inspect messages
• retrieve data
• examine records

or:

Act

• create something
• update something
• send something
• schedule something
• change a record

Treat those differently.

Reading your calendar to find availability is one risk level.

Sending a message to a customer is another.

Human review remains useful whenever an AI action creates financial, reputational, contractual, or customer-facing consequences.

---

Step 7 — Let Several AI Agents Work at the Same Time

One of the biggest gains from agent workflows is not smarter prose.

It is parallel work.

Suppose you ask one AI:

Find whether there is a business opportunity in AI video editing.

It might research for several minutes and return one interpretation.

Instead, you could run four independent lines of inquiry:

Agent A: What are customers complaining about?

Agent B: What products already exist?

Agent C: What are people searching for?

Agent D: Where are businesses already spending money?

Then the director compares the evidence.

OpenAI explicitly positions Codex as a multi-agent command center that can run agents in separate threads and parallel workflows. (OpenAI)

Claude Agent Teams take this idea further in their own direction. Anthropic distinguishes teams from ordinary subagents and packages agent collaboration visibly in Claude Code. Claude plugins can bundle Skills, hooks, subagents, and MCP servers together. (Claude Platform Docs)

This is one area where I give Claude credit.

Anthropic makes the architecture easy to see.

OpenAI has many similar pieces, but they can appear across ChatGPT, Work, Skills, Plugins, Projects, Codex, and developer tools.

That makes ChatGPT look less agentic than it really is.

TRY THIS IN CHATGPT

Research this business idea from four independent perspectives. Do not let the agents converge early.

Agent 1: customer pain.

Agent 2: competition.

Agent 3: search demand.

Agent 4: monetization.

Once the independent reports are complete, compare them. Identify conclusions supported by multiple independent lines of evidence. Highlight disagreements rather than hiding them.

That last sentence matters.

Do not ask AI teams only to agree.

Ask them to find where they disagree.

---

Step 8 — Put Your ChatGPT Workflows on a Schedule

This is where AI starts moving from something you remember to use into something that remembers for you.

ChatGPT Scheduled Tasks can handle reminders, recurring jobs, briefings, and monitoring. OpenAI distinguishes these from Codex Automations, which run workflows inside Codex. (OpenAI Help Center)

Consider this schedule:

Monday — Opportunity Research

Find five rising AI business topics that show growing interest, commercial intent, and meaningful customer problems.

Wednesday — Competitor Watch

Check the competitors we are tracking and report meaningful product, pricing, positioning, or marketing changes.

Friday — Content Opportunity Report

Review this week’s findings and identify the three strongest article or video opportunities.

Now your system has rhythm.

ChatGPT Work extends this idea into longer jobs. OpenAI says Work can gather context from files and plugins, operate across workflows, and create reviewable outputs such as documents, presentations, spreadsheets, Sites, and other finished work. (OpenAI Developers)

TRY THIS IN CHATGPT

Every Monday, research meaningful new developments in [your industry]. Rank the five developments most likely to create a business, content, or customer opportunity. Ignore routine announcements. Notify me with a concise report containing what changed, why it matters, and what action may be worth considering.

Now you have something much closer to an ongoing research analyst than a chatbot you occasionally remember to question.

---

Step 9 — Give the Entire AI Team One Goal

This is where the pieces come together.

Most people manage AI at the task level:

Write this paragraph.

Research this keyword.

Fix this headline.

Agentic work moves one level higher:

Here is the goal. Decide what work is required.

For example:

Find a promising AI business topic and turn it into a publishable content package.

The director may decide that this requires:

1. trend research
2. market validation
3. competitor analysis
4. SEO research
5. source verification
6. article planning
7. drafting
8. fact checking
9. visual planning
10. distribution planning

Some of those jobs can happen at the same time.

Others depend on earlier findings.

That is orchestration.

The Master Prompt

You are the director of an AI marketing department.

Objective: Find one promising AI-related business topic and turn it into a publication-ready content opportunity.

Delegate research to specialist agents where supported.

Research market interest, customer pain, competitor activity, search intent, and monetization potential independently.

Reject ideas supported mainly by hype.

Select the strongest opportunity using evidence from multiple research paths.

Create:

• the core angle
• target reader
• search intent
• headline options
• article outline
• source plan
• differentiation strategy
• visual plan
• distribution plan

Before finalizing, run a separate verification pass that checks claims, dates, software capabilities, and unsupported assumptions.

Show me the final recommendation and the evidence that led to it.

This prompt is not magic.

The structure is what matters.

One objective.

Separate specialists.

Independent research.

Synthesis.

Verification.

Human review.

---

Build This: A Seven-Person AI Content Company Inside ChatGPT

Here is the complete system:

Marketing Director
Controls objectives and delegation.

Trend Researcher
Finds what is changing.

Keyword Researcher
Finds what people want.

Competitor Analyst
Finds what already exists.

Article Writer
Turns evidence into useful content.

Fact Checker
Challenges unsupported claims.

Visual Director
Plans imagery and explanatory visuals.

Distribution Strategist
Finds ways to turn one piece into multiple channels.

Now consider what happens when you combine that organization with:

Project
Keeps the business context.

Project Instructions or AGENTS.md
Keeps the operating rules.

Skills
Keep repeatable procedures.

Subagents
Perform specialist work.

Apps/MCP
Provide access to real systems.

Scheduled Tasks
Trigger recurring work.

Work/Codex
Handle longer, multi-step execution.

At that point, you are no longer using ChatGPT merely as a question-answering box.

You are building reusable operating systems around AI.

---

7 Things You Probably Didn’t Know ChatGPT Can Do

1. ChatGPT Can Use Reusable Skills

Skills can package recurring procedures instead of forcing you to rewrite giant prompts. (OpenAI Help Center)

2. ChatGPT Work and Codex Can Use Subagents

Specialized agents can be assigned separate parts of a job and their findings can be collected into a final response. (OpenAI Developers)

3. Codex Can Run Multiple Agents in Parallel

The Codex app was built around multiple concurrent agent threads and parallel workflows. (OpenAI)

4. Codex Can Read Permanent AGENTS.md Instructions

You can define project rules that Codex reads before starting work. (OpenAI Developers)

5. ChatGPT Can Work From Connected Apps

Plugins and Apps can connect workflows to outside data and actions. (OpenAI Help Center)

6. ChatGPT Can Perform Scheduled Work

Scheduled Tasks can run one-time jobs, recurring jobs, briefings, and monitoring. (OpenAI Help Center)

7. ChatGPT Work Can Turn a Goal Into Finished Deliverables

Work can operate across files and connected context and produce documents, spreadsheets, presentations, Sites, and other finished outputs. (OpenAI)

That last capability is especially interesting.

The future may not be:

“Ask AI better questions.”

It may be:

“Give AI better systems.”

---

Where Claude Still Has an Edge

This article would be misleading if I pretended the products were identical.

They are not.

Claude Makes Subagent Architecture Very Visible

Claude Code has made subagents a central part of its workflow. Users can create specialized agents and reason about them as separate workers. (Claude Platform Docs)

Claude Agent Teams Are More Explicit

Anthropic has pushed multi-agent collaboration into a clear concept that is easy to demonstrate.

OpenAI can accomplish parallel and multi-agent work through several surfaces, but the mental model can feel less unified.

Claude Hooks Are Different

Claude Hooks let users trigger deterministic commands at defined stages of the Claude Code lifecycle. Anthropic recommends hooks for automation and rule enforcement where deterministic behavior matters. (Claude Platform Docs)

There is no single normal-ChatGPT button that maps perfectly onto that.

Claude Plugins Package Agent Components Neatly

Anthropic plugins can bundle Skills, hooks, subagents, and MCP servers into one installable unit. (Claude Platform Docs)

That packaging is genuinely attractive.

Giving Claude credit here makes the comparison more useful.

---

Where ChatGPT May Be Easier

ChatGPT has a different advantage:

You can begin with almost none of this terminology.

Create a Project.

Give it instructions.

Upload your material.

Ask it to research.

Use a Skill.

Connect an App.

Schedule a task.

Move into Work when you need a larger multi-step job.

Move into Codex when you need deeper agent workflows.

OpenAI says ChatGPT Work is available across desktop and, for several paid plans, web and mobile, and it is intended to turn larger goals into finished work. (OpenAI)

Codex is spreading beyond software development too. OpenAI has described nontechnical teams using Codex for internal apps, dashboards, executive materials, research, creative briefs, and other knowledge-work tasks. (OpenAI)

That makes the dividing line between “chatbot” and “agent platform” much harder to see.

---

The Biggest Mistake: Comparing Feature Names Instead of Outcomes

Claude has Skills.

ChatGPT has Skills.

Claude has subagents.

OpenAI has subagents and multi-agent systems.

Claude has CLAUDE.md.

Codex has AGENTS.md.

Claude has MCP.

OpenAI supports external tool connections and MCP-based workflows.

Claude has agent tooling.

OpenAI has ChatGPT Work, Codex, and the Agents SDK. OpenAI’s multi-agent APIs can coordinate subagents in parallel and synthesize their work. (OpenAI Developers)

But chasing terminology misses the bigger lesson.

Ask yourself five questions:

What information should the AI remember?

What procedure should it reuse?

What jobs should be separated?

What tools should it access?

What should happen automatically?

Those five questions matter more than whether a YouTube creator calls something a Skill, agent, subagent, plugin, command, or workflow.

---

Which Should You Use: Claude or ChatGPT?

Use the system that fits the work.

Claude may feel particularly natural if you love Claude Code, terminal-centered workflows, visible subagent structures, Anthropic’s plugin model, or Claude’s agent-team approach.

ChatGPT may feel natural if your work already lives in ChatGPT Projects, connected Apps, Skills, Work, Scheduled Tasks, Codex, or OpenAI’s broader ecosystem.

There is no prize for loyalty to a software interface.

There is a prize for creating a workflow that saves time and produces better work.

And there is no reason you cannot use both.

---

The Real Opportunity Is Not Claude vs. ChatGPT

The most interesting lesson I found from researching Claude was not that I needed to leave ChatGPT.

It was that many of us are still using AI at a much lower level than the tools now allow.

We learned to prompt.

Then we learned to write better prompts.

Now the next step is learning to create:

context

instructions

Skills

specialists

tools

delegation

verification

automation

That changes the question from:

“What should I ask ChatGPT?”

to:

“What system should I build so I do not have to keep asking the same thing?”

Start small.

Create one Project.

Give it one operating manual.

Turn one repeated prompt into a Skill.

Delegate one research job to specialists.

Schedule one recurring task.

Then watch what happens.

The people getting the most from AI over the next few years may not be the people who discover the cleverest prompts.

They may be the people who learn how to turn AI into repeatable systems.

And you do not have to switch to Claude to start doing that.

You can start building it inside ChatGPT today.