How to Use ChatGPT Work Better Than 99% of People (Full Guide)

Technological concepts & product features covered

“ChatGPT Work” as an AI agent (vs. regular chat)

• Regular ChatGPT is positioned as a thinking partner: it returns answers, but the user still does the work.
• ChatGPT Work (and Codex) are described as AI agents that can:
  • Delegate a multi-step task
  • Make a plan
  • Work through steps until completion

Same underlying agent across interfaces (Work vs Codex)

The speaker claims both Work and Codex use the same agent, presented through different product interfaces:

• ChatGPT Work → knowledge work
• Codex → code-focused work

Key use case for non-technical teams

ChatGPT Work is recommended for teams like finance, legal, HR, marketing, and sales, especially when tasks require multi-step operations across tools and documents.

---

Tutorial/guide structure: “8 steps” to become a power user

1) Setup & access method differences (desktop vs web/mobile)

• Download the ChatGPT desktop app for the full feature set, including access to Work across plan tiers.
• The web version is said to have reduced feature coverage and prompts users to open the desktop app.
• The desktop app includes Work in two modes:
  • Local mode (on your computer): the agent can access your local files
  • Cloud mode: runs in the cloud but doesn’t access local files, and (per the speaker) lacks the local in-app browser
• The main advantage highlighted: Local mode + the inbuilt browser

2) Projects for file-based workflows

• Put each task into a dedicated project.
• These projects differ from regular ChatGPT “projects” and support local folders as source folders.
• Two file-provision approaches are contrasted:
  • Attaching via “plus/attach files and folders” is shown as less capable (only the built-in browser visible)
  • Correct project source-folder setup unlocks:
    • a Files panel
    • a Browser panel
    • direct agent visibility/processing of files
• Projects can have multiple source folders, but only one primary folder.
• Example workflow:
  • process receipt images
  • extract fields
  • generate Excel
  • validate
  • move/organize processed receipts into month subfolders

3) Agent planning + steering

• Work tasks take longer because they run as agents.
• Agent-generated plan steps are shown in the UI.
• If the user disagrees, they can:
  • Stop the agent and adjust
  • Send a follow-up prompt to steer the agent

4) Confidence scoring + debugging

• Example deliverable includes a confidence score per extracted record.
• The UI allows reviewing notes for items needing manual review.
• There’s also an interface to inspect detailed working steps for error/debugging.

5) agents.md for persistent instructions & context

• agents.md is described as a persistent instruction/context file that Work reads at the start of a project session.
• Intended to:
  • avoid repeating instructions
  • speed up tool usage by supplying structure/context (e.g., a Notion database location)
• Demonstrated speed improvement:
  • without agents.md: search through tools/Notion pages (~43 seconds)
  • with agents.md: returns the correct link in a few seconds

6) Plugins (apps + skills)

Plugins are described as more than “connections”:

• A plugin is an installable package containing:
  • Apps: where to connect/act
  • Skills: how to perform repeatable workflows
• Examples:
  • Gmail plugin: includes both app + skills (search, compose, send guidance)
  • Presentations plugin: described as skill-only (no app connection)
  • Data analytics plugin:
    • installs first
    • then prompts users to connect multiple relevant apps (e.g., Slack/Teams/Gmail)
    • enables skills like dashboards/reports/visualization

7) Skills (custom reusable workflows)

• The video differentiates:
  • skills inside plugins
  • user-created skills
• Skills are reusable, step-by-step workflow instructions (often referencing templates) that can be enabled “on demand.”
• Demonstrated outcome:
  • Without a dedicated skill: email creation ~8+ minutes, more generic styling
  • With a dedicated “create marketing email” skill: ~1:50, matching a specific template/style (logos, colors, buttons, footer links)
• Custom skill creation workflow:
  1. do the task once step-by-step with Work until results are correct
  2. refine instructions and results
  3. use an inbuilt skill creator (via “/skill creator”) to generate/install the skill
  4. test and iterate; update skills again using the skill creator
• Skills can reference other skills (example: blog feature image workflow uses an image-generation skill)

8) Two “power features”: Sites and Browser use

Sites

• Lets you build and deploy interactive websites/dashboards from Work.
• Demonstrated: a receipt inbox web app where a phone takes photos and receipts get processed.
• Includes:
  • a hosted site on chat.openai.com with login-gated access
  • API key/settings management to avoid exposing keys in source code
  • example output: an extracted receipt record (gross amount, currency, category, description)
• Sharing options:
  • share with anyone in the workspace or specific individuals
  • warned as risky: “anyone on the internet”
• Sites can also be created via “fork/spin off” flows from prior context.

Inbuilt browser

• Lets Work open and interact with web pages like a normal browser.

• Enables automation by prompting Work to:

  • read a page directly (not rely on snippets/search)
  • extract structured data into a file inside the current project (example: convert Claude release notes into a CSV with date/title/description/links)

• Example scheduled monitoring task:

  • check a competitor’s release notes 4x/day (every 6 hours)
  • compare against an existing CSV
  • if new items appear:
    • append to CSV
    • send an email using a connected Gmail connector
• Browser settings include an inbuilt password manager, enabling automation on sites that require login even without a direct API.

---

Key analysis/product positioning and outcomes

• The video’s main argument: agentic Work + local files + projects + agents.md + skills/plugins + sites/browser automation enables more end-to-end execution than typical ChatGPT prompting.
• Repeated emphasis that local file access is critical for real work completion (not just answers).
• Emphasis on automation and operationalization:
  • scheduled monitoring
  • template-based output
  • deployment of interactive tools

---

Main speakers/sources

• Main speaker: video presenter/host (no specific name provided in the subtitles)
• Referenced systems/products: ChatGPT Work, Codex, Notion, Gmail, customer.io, HubSpot (example), BigQuery, Anthropic/Claude