ChatGPT Work: 14 Capabilities

Credit — Riley Brown

Source video by Riley Brown (@rileybrown) — Agent Native / agents for business; podcast @agentnative_.

Original post: x.com/i/status/2085761931403575356 (7 Aug 2026) — “Learn 99% of ChatGPT Work in 61 minutes.” ~63k views at capture; heavy bookmark signal.

Framing from the video: if you use ChatGPT, Claude, Claude Cowork, or OpenClaw but still don’t know which agent to use for business and how — this is the map for ChatGPT Work (OpenAI’s agent surface: like Cowork, but Riley’s bet is it’s stronger and more usable across mobile / web / desktop).

Capability list, demos, and workflow opinions below are his. Our extras: links into existing Work / Codex super app / Voice notes and where this fits multi-company / HEOS-style ops.

Timestamps (from the post)

What ChatGPT Work is

In ChatGPT, flip the Work toggle (web and iOS). You’re no longer in plain chat — you’re in an agent that runs on a cloud computer, produces real files (pptx, docx, pdf, html, spreadsheets, sites), and can use plugins / skills for hours of multi-step work.

Riley’s one-liner: GPT Work ≈ more accessible Codex — same “build anything” energy as the Codex desktop super app, but available on web + iOS + desktop (cloud), not only a local coding-flavored shell.

How it got here (timeline he tells)

1. 2022 ChatGPT — first mass-accessible AI chatbot → nearly a billion weekly actives over ~4 years.
2. Codex desktop — OpenAI’s reaction to Claude Code / Claude Cowork: a super app for apps, decks, video, “anything,” not only coding. Name still felt developer-only.
3. Merge ChatGPT + Codex surfaces into ChatGPT → ChatGPT Work (~6 months after Codex): accessible agent path on web, desktop, and iOS.

Closest alternatives on this site: Claude Cowork vs Eigent, OpenWork, OpenClaw — same “do the work” category, different host and trust model.

The 14 capabilities (business summary)

1. Polished reports & documents

What: Agent uses a cloud computer to research and ship real files — .pptx, .docx, .pdf, inline .html — editable in the side panel.

Demo pattern: “McKinsey-style AI outlook for rest of 2026, heavy research + graphics.” ~13–14 minutes of work, 87 sites searched, 19-slide deck. Edit by blabbing slide-by-slide feedback (“fix graphic on slide 1, kill the pills on slide 2, recolor slide 5…”) — ~10 more minutes, coworker-style iteration.

Business use: Board decks, client reports, internal strategy docs without PowerPoint jockeys. Mobile hack: open the artifact on iOS and voice-edit → new PDF/doc revision in minutes.

2. Plugins (official app connections)

What: Official integrations (not the same as skills). Browse/install from Work → plugins. Suggestions in the UI can pull from calendar, email, Slack, etc. once connected.

His stack examples: ClickUp (team), Convex (DB), built-in Documents, FAL (image/video models), GitHub, Gmail, Calendar, Google Drive, Hostinger, HeyGen Hyperframes, Notion, Canva (one-tap OAuth).

Business use: “Has Emily emailed?” → draft reply → update Notion → schedule calendar invite with Emily → optional one-shot automation if he doesn’t follow through. No @-mention required every time once installed; model picks the plugin. Install works from mobile settings too.

Plugins bundle skills (e.g. Gmail read/send). Some plugins expose dozens of skills; some one.

3. Blocks

What: Editable in-chat artifacts you can full-screen, type into, or ask the agent to revise — with version history (back/forward) and suggestion chips (e.g. stronger hook).

Types he shows: writing blocks (podcast intro), Mermaid mind maps / flowcharts / sequence diagrams, email draft blocks, data/technical, media, file & app blocks. Ask Work: “list all block types” to inventory.

Business use: Tight collaboration on one intro paragraph, architecture diagrams for a process, sequence maps of how Work + plugins + skills produce an output — then lift those into a site (capability 4).

4. Websites, apps & hosting

What: Build + deploy shareable mini-apps via ChatGPT Sites (chatgpt.site-style internal hosting), then promote to the real internet with plugins (demo: Vercel live URL; optional Namecheap + custom domain).

Business use: Public “what blocks/charts can Work render?” gallery, internal tools, landing pages — from the same thread as research/diagrams. Same prompts work from phone.

5. Image generate & edit (GPT Image 2)

What: Built-in image model for gen + surgical edits: swap text/logo, keep rest; multi-variations; inline edit while generating; pin comments on regions (“more beard,” “blue eyes,” gradient tweak).

Business use: Thumbnail iterations → attach four variants in Gmail to a coworker for pick — plugins + images in one flow.

6. Branching chats

What: /branch chat forks context into a parallel thread; rename; /pin to pin. One parent exploration → one branch continues the site, another becomes a PDF deck.

Business use: One task per branch. Stops the “one chat does everything and context rots” pattern.

7. Desktop app — cloud vs local

ChatGPT + Codex apps merged; desktop Work has a critical switch:

Rule of thumb: Cloud for multi-device + always-on email loops. Local when you need disk + local skills but fear Codex full control. His split: phone Work ~80%; desktop coding-heavy → Codex more than Work.

8. Voice mode

What: “Start new voice chat” on Work — full conversational control of plugins (Notion, Gmail, browser). Live demo: open the right Notion page in Codex browser, fix wrong title, rewrite page headings, add bullets under a section while talking.

Business use: Sponsorship lead table from inbox; draft reply with draft skill; open in browser for human send. Anything plugins can do, voice can drive.

Related deep dive (different creator, same era): Alex Finn — ChatGPT Voice advanced tips.

9. Remote voice mode (phone → desktop)

What: iOS Remote tab pairs phone voice to desktop ChatGPT. Master voice thread can spin up new GPT Work sessions (and Codex sessions) by name while you walk.

Daily workflow: Arrive, plug in laptop, walk 30 minutes with phone. Process email by voice; for each task, “spin up a new Work session…” Come back to 4–5 threads mid-flight or ready drafts.

Business use: Chief-of-staff orchestration without sitting at the desk. Complements local browser prep (capability 10).

10. In-app browser (underrated)

What: Full browser inside the desktop surface (new tab / Cmd+T). Sign into accounts (import cookies/passwords). Local Work can drive the UI (post to X, open Gmail drafts, Google Docs).

Business use: Task-scoped tab sets: one Work thread + the tabs that matter for that task. When spinning sessions from the phone, ask it to open the two tabs you’ll need when you sit down. Habit: open external links “in browser” so work stays in one app.

11. Skills

What: Instruction files (skill.md-style) the agent loads for repeatable workflows — YouTube thumbnail skill, Notion video DB skill, etc.

Critical split:

• Local skills (created in Codex / local Work) → available on local Work, not on cloud phone/web Work.
• Cloud skills (created in cloud Work / web / phone) → available on all cloud surfaces.

Want a walk-and-talk skill on iOS? Create it in the cloud. Riley’s product ask: OpenAI should unify skill stores. Skills culture on this site: Claude skills, Trae skills.

12. Scheduled automations

What: Work that runs in the future — one-shot or recurring checks while you’re offline.

Pattern: Send email Friday → starting Monday 9am every hour until next week, check for reply → append response to Notion → email you on vacation address when done. Same pattern for “if I don’t finish, send her the link in 6 hours.”

Must be cloud Work (web / cloud desktop / phone). Local / pure Codex schedules die when the machine sleeps. Mind shift: the agent is not only “do this now” — it’s a lightweight ops runner for response routing.

13. Spreadsheets

What: Frontier models are “almost scary” at multi-sheet research workbooks with charts — e.g. deep MSFT / GOOG / AAPL comparison, sources tab, then dashboard cover sheet + convert to Google Sheets via Drive plugin.

Business use: Competitive financial packs, ops dashboards from internal docs + web research, shareable Sheets without Excel heroics.

14. Multi-agent workspace

What: Several cloud Work threads at once (side panel, compact multi-view). Double-click to full-screen a task, snap back. Mix Work tabs and Codex tabs. Agents take longer now — parallel is the default.

Business use: Email triage in one thread, podcast intro in another, chart gen in a third — hop without waiting for each long job to finish.

Mental model: two skill planes

Remote voice bridges them: phone is the orchestrator; desktop is the factory floor that spins workers.

Where this fits our stack

Quick start checklist

1. Enable Work toggle on web and install iOS app Work toggle.
2. Connect the boring plugins first: Gmail, Calendar, Drive, Notion (or your stack).
3. Ship one artifact end-to-end: research deck or spreadsheet — practice side-panel edit loops.
4. Create one cloud skill you need on walks; keep Codex-only skills local on purpose.
5. Set one cloud automation: “check for reply → notify.”
6. Pair phone remote voice to desktop; walk one email session spinning Work threads.
7. Decide cloud vs local per task before you start (sync vs disk).

Related on this site

• ChatGPT Work: summary & action guide (earlier field notes)
• ChatGPT + Codex super app (part 2)
• ChatGPT Voice: 8 advanced tips (Alex Finn)
• Alex Finn: reverse prompting
• Claude Cowork vs Eigent
• OpenClaw · OpenWork
• CLI for agents · Multi-agent harness map