ChatGPT Work Complete Guide: 14 Capabilities for Real Business

Direct Answer

ChatGPT Work is most useful when you stop treating it as a smarter reply box and start treating it as a place to delegate a defined job. Riley Brown's hour-long tutorial demonstrates 14 capabilities across presentations, plugins, visual blocks, websites, images, branched chats, desktop work, Voice, Remote, Browser, skills, schedules, spreadsheets, and parallel workspaces. The real lesson is not that every feature should be used at once. It is that one task can now move from source material to a finished, reviewable deliverable without a chain of disconnected AI tools.

The strongest setup has five layers: context, tools, execution, review, and reuse. Give the task reliable source material, connect only the tools it needs, define the output and completion checks, review sensitive actions, then save the successful procedure as a skill or schedule. That is the difference between an impressive demo and a business system.

Watch the Tutorial

Video credit: Riley Brown. Watch the original video on YouTube. The demonstrations and workflow examples come from Riley's tutorial. Platform availability, product terminology, and local-versus-cloud boundaries below were checked against official OpenAI learning documentation on 9 August 2026.

What ChatGPT Work Actually Is

Riley describes Work as "Codex in the cloud." That is an intuitive starting point, but the official boundary is more useful: Chat is for an answer; Work is for an outcome. Use Chat for an explanation, a short draft, or brainstorming. Use Work when you want a brief, deck, analysis, recurring update, website, spreadsheet, or other artifact that can be reviewed against a definition of done.

Codex overlaps with that task model, especially in the desktop app, but it reaches deeper into implementation. Local Codex can operate in a selected repository or folder, run commands, edit files, inspect diffs, execute tests, and coordinate development agents. Cloud Work is portable across devices and can use uploaded files and approved connections, but it does not automatically inherit every file, login, or local skill from your computer.

The Surface Map: Web, Desktop, Mobile, CLI, and IDE

The word "ChatGPT" now covers several surfaces that share context but do not expose identical capabilities. Choose the surface from the job's dependencies, not from habit.

This distinction changes reliability. A cloud schedule can run while your laptop is closed, but it cannot inspect a new file sitting in a local folder. A local schedule can use that folder, but only while the computer and app are available. A Remote session lets your phone direct work on a paired computer; it does not turn every local task into a permanently hosted cloud worker.

The 14 ChatGPT Work Capabilities

Riley's description accidentally skips number five in the written chapter list. The transcript fills it in: image generation and editing. Here is the complete map, with the practical job and the control each capability needs.

Create: Finished Artifacts, Not Longer Answers

1. Presentations

In the demonstration, Work researches dozens of sources and produces a 19-slide, consulting-style presentation, then revises it through Voice and exports multiple file types. The transferable workflow is stronger than the exact numbers: define the audience, argument, evidence standard, slide count, visual system, notes, and acceptance checks before generation.

Do not ask only for "a beautiful deck." Ask for one sentence per slide that advances an argument, a source note for every quantitative claim, an appendix for excluded evidence, and a final pass that removes decorative repetition. A model can fill 40 slides. The harder job is deciding which 12 deserve to exist.

3. Blocks and visualizations

The video uses "Blocks" as shorthand for editable content blocks and inline visual outputs such as Mermaid mind maps, flowcharts, sequence diagrams, timelines, and user journeys. Current OpenAI documentation describes the formal interactive capability as Visualizations: charts, maps, diagrams, calculators, simulations, and other rendered tools. The exact UI can change; the durable idea is to turn an explanation into an inspectable object.

4. Sites and 5. Images

Sites can turn the research itself into a hosted report, interactive calculator, lightweight app, or campaign page. Contrary to a common shortcut, a ChatGPT Site is a real hosted deployment. Vercel or another host becomes useful when you need a separate repository, custom delivery pipeline, environment variables, or infrastructure ownership.

Image generation completes the artifact loop. Riley edits a reference image, creates variants, and sends the result through a connected tool. Keep a simple approval rubric: required dimensions, factual text, logo treatment, product fidelity, representation, licensing, and a final human check at the actual placement size.

13. Spreadsheets

A spreadsheet is useful when the reasoning needs rows, formulas, filters, charts, or reconciliation. Ask Work to keep a source sheet, an assumptions sheet, a calculation sheet, and an executive dashboard separate. A polished chart can still rest on a broken cell reference, so require formula inspection, units, missing-data flags, and tie-outs against the source totals.

Connect: Plugins, Tools, and Business Data

Riley demonstrates connected work across Gmail, Calendar, Drive, Notion, GitHub, ClickUp, Vercel, and other services. The official vocabulary matters:

• Skill: a reusable workflow with instructions, templates, examples, schemas, and review criteria.
• Connector: a controlled connection to external tools or data, commonly backed by MCP.
• Plugin: an installable bundle that can contain skills, connectors, or both.

Connections create leverage because the agent can finish the work where it lives. They also create blast radius. A Gmail reader, calendar writer, production deployer, and finance database should not share the same default approval policy. Start read-only, enable one narrow write action, keep logs, and require a confirmation that states the target, change, and rollback path.

Coordinate: Branches, Desktop, Voice, Remote, and Browser

6. Branching chats

Branching is useful when one body of context should produce two independent deliverables. Riley branches a website conversation into a presentation task. That is cleaner than asking one thread to keep changing roles. Name branches by output, pin the active ones, and end each branch with a decision memo so the useful result can return to the main project.

7. Cloud or local

Choose cloud when the job depends on uploaded files, connected cloud tools, portability, or an always-on schedule. Choose local when it depends on files, apps, browsers, development environments, or tools on the computer. The local option is not automatically "better" because it has more access. It is appropriate only when the task needs that access.

8. Voice and 9. Remote Voice

Voice is best as a direction layer: state the outcome, answer clarifying questions, inspect progress, and correct assumptions without stopping to write a perfect prompt. Full Voice behavior requires starting the task in Voice mode; a task begun as text may expose dictation instead. Remote extends that control to a paired iOS device while the work continues on the connected computer.

A useful voice command contains four parts: goal, context, authority, and next checkpoint. For example: "Review the launch folder, draft the three missing assets, do not send or publish anything, and call me back when the evidence table is ready." That is more dependable than an open-ended instruction to "finish the launch."

10. In-app Browser

Browser work is valuable for websites without an API, visual inspection, research, form preparation, and repetitive web steps. Treat every page as untrusted input. Ask the agent to stop before purchases, submissions, deletion, public posts, credential changes, or any action that cannot be reversed. Use the Chrome extension only when the task truly requires an existing Chrome profile or tab.

Reuse and Scale: Skills, Schedules, and Parallel Work

11. Skills

A skill should encode the judgment that made the workflow succeed, not just preserve the original prompt. Include the trigger, accepted inputs, ordered steps, tool permissions, output schema, examples, failure cases, and definition of done. Keep local and cloud availability in mind: a skill built around a local executable cannot become portable merely because its instructions sync.

12. Scheduled automations

Riley's example sends an email, checks periodically for a reply, routes the response into Notion, and alerts the right person. That pattern is useful, but polling every hour forever is not a complete automation. Add a start time, end time, success condition, timeout, deduplication key, escalation owner, and rule for partial failure.

14. Multi-agent workspace

The video demonstrates four cloud Work windows running in parallel. That is parallel task management, which is already valuable. Official subagents go one step further: a lead task delegates independent jobs to specialized agents and gathers their results. Do not confuse more windows with orchestration. Reliable multi-agent work still needs task ownership, isolated write targets, shared state, acceptance tests, and one agent or human responsible for integration.

The Five-Layer Business Operating System

1. Context: place durable files, instructions, decisions, and source boundaries in a project. Keep one task per chat.
2. Tools: connect only the plugins, browser profile, local folder, or APIs required for that job.
3. Execution: state the outcome, constraints, format, budget, deadline, and definition of done. Use branches or subagents for independent paths.
4. Review: require evidence, tests, reconciled totals, visual inspection, and explicit approval before sensitive writes.
5. Reuse: turn a proven process into a skill. Schedule it only after repeated manual runs produce accepted outcomes.

The order matters. Automating before defining context and review creates faster confusion. A small, observable workflow with strong checks will outperform a heroic instruction that asks one agent to run the entire company.

Four Copy-Ready Work Prompts

1. Source-backed presentation

Goal
Create a 12-slide decision deck for [audience] about [decision].

Sources
- Use only the attached files and approved web sources.
- Put a source note beside every numerical claim.
- Flag conflicts instead of silently choosing one source.

Structure
1. Executive answer
2. Evidence
3. Options and tradeoffs
4. Recommendation
5. 30-day action plan
6. Appendix with rejected or uncertain evidence

Design
Use the attached brand guide. One idea per slide. No decorative charts.

Definition of done
Every claim is traceable, the recommendation follows from the evidence,
and the deck has been checked for unsupported numbers and repeated slides.

2. Site with a review gate

Build a responsive internal site from the approved report.

Include
- Executive summary
- Filterable evidence table
- Timeline
- Risk register
- Downloadable source list

Rules
- Do not invent data or testimonials.
- Keep all external links labeled and working.
- Meet basic keyboard, contrast, and mobile layout requirements.
- Save a reviewable version first. Do not deploy until I approve it.

Return
The preview, a list of assumptions, and a deployment checklist.

3. Scheduled weekly operating brief

Every Friday at 15:00 Europe/Lisbon, create a weekly operating brief.

Read-only sources
- [Project tracker]
- [Approved Slack channels]
- [Analytics source]

Deliverable
- Five-line executive summary
- Work completed with evidence links
- Blockers, owner, and age
- Metric changes with previous-period comparison
- Decisions needed next week

Controls
- Do not post, send, or modify source systems.
- Mark unverifiable claims as unverified.
- Stop and notify me after two consecutive source failures.
- Save the brief as a draft for human review.

4. Parallel research pack

Goal
Prepare a decision pack on [topic].

Delegate independent work
- Agent A: market evidence and primary sources
- Agent B: customer pain points and counterexamples
- Agent C: technical feasibility and dependencies
- Agent D: legal, security, and operational risks

Isolation
Each agent writes to its own file and cites every important claim.
No agent edits another agent's output.

Integration
A final reviewer identifies conflicts, missing evidence, and assumptions.
Do not average disagreements. Present them clearly.

Definition of done
The final memo includes recommendation, confidence, open questions,
decision criteria, and links to every source file.

A Practical Permission Ladder

A Seven-Day Setup for a Small Team

1. Day 1: choose one recurring job that takes 30 to 120 minutes and has a clear accepted output.
2. Day 2: collect the minimum trusted context and define what the agent must never infer.
3. Day 3: run the task in cloud Work with no write permissions. Measure corrections and missing evidence.
4. Day 4: test the same task locally only if it needs local files, apps, Browser, or terminal tools.
5. Day 5: write an acceptance checklist and package the successful procedure as a skill.
6. Day 6: add one controlled plugin or automation with a stop condition and human review.
7. Day 7: compare time, accepted quality, cost, failure rate, and review burden. Keep only what improved the result.

A mature setup is not measured by how many plugins appear in the directory. It is measured by accepted outcomes per hour of human attention.

Video Chapters

1. 0:00 - Introduction
2. 3:18 - Presentations
3. 9:16 - Plugins
4. 17:36 - Blocks and visualizations
5. 23:34 - Sites, apps, and hosting
6. 26:04 - Image generation and editing
7. 30:09 - Branching chats
8. 31:51 - Desktop app: cloud versus local
9. 37:01 - Voice Mode
10. 41:36 - Remote Voice Mode
11. 45:52 - In-app Browser
12. 48:50 - Skills
13. 51:45 - Scheduled automations
14. 54:37 - Spreadsheets
15. 56:58 - Multi-agent workspace
16. 59:19 - Final review

Bottom Line

Riley Brown's tutorial is valuable because it shows ChatGPT Work as a connected work surface rather than a collection of isolated features. Presentations, Sites, images, visualizations, and spreadsheets create outputs. Plugins and Browser bring in tools and data. Branches, Voice, Remote, desktop modes, and parallel tasks coordinate execution. Skills and schedules preserve what works.

The durable advantage is not knowing 14 buttons. It is knowing how to turn a repeated job into a controlled loop: trusted context, narrow tools, explicit outcome, observable execution, human review, and reusable procedure. Build one loop that produces accepted work every week. Then add the next one.

Sources

• Riley Brown: Learn 99% of ChatGPT Work in 61 Minutes (Codex for Work)
• Riley Brown on YouTube
• OpenAI: Get started with ChatGPT Work
• OpenAI: Skills and plugins
• OpenAI: Plugins across ChatGPT and Codex
• OpenAI: Projects and chats
• OpenAI: Sites
• OpenAI: Browser
• OpenAI: Voice
• OpenAI: Remote
• OpenAI: Visualizations
• OpenAI: Scheduled tasks
• OpenAI: Long-running work
• OpenAI: Subagents