How To Use ChatGPT Work: The Complete Beginner's Guide (2026)

Technological concepts & product features highlighted

ChatGPT Work vs. Codex (and model selection)

• The video frames ChatGPT Work as a “place” for everyday work tasks (not necessarily coding).
• Codex is described as strongest for software engineering / editing files / applying changes to code or documents.
• Models like “Astra” / “Soul” / “Terra” / “Luna” are discussed as interchangeable “agents” with different strength/cost/effort tradeoffs for:
  • demanding tasks
  • lightweight tasks
  • repetitive tasks
• Mentions accessibility and “Astra” availability differences across interfaces (Work/Codex vs regular ChatGPT).

Cloud vs Local execution (keeping tasks running after you leave)

• Core idea: tasks can continue after you close your laptop if you run them in cloud mode.
• Local mode can run in the background, but requires:
  • the computer stays on
  • the relevant application stays launched
  • the machine does not sleep
• Key pitfall: cloud tasks don’t automatically access your local desktop files—you must upload or place files in an accessible location before leaving.
• Guidance examples:
  • “tomorrow morning with laptop off” ⇒ cloud
  • “Mac folder-specific work” ⇒ local

Voice input and screen context

• Two voice mechanisms:
  • Dictation: transcribes language to text for review before sending
  • Voice chat: more conversational, with limited tool access
• Mentions a Mac setting for context/screen awareness, potentially including off-screen text (e.g., scroll content).
• Mentions hotkeys for voice and using voice to add missing context without restarting.

Surgical document editing (Word/Google Docs/DOC)

• Emphasizes editing existing documents without rewriting everything.
• “Be specific” about:
  • what to change
  • what to preserve (headers, links, sections, comments under paragraphs)
• Uses “copy then verify” as a beginner-safe workflow:
  • request a copy first to preserve formatting and links
  • later switch to direct editing once trusted
• Strong advice: verify the actual file after changes (formatting, tables, overflow issues, correctness).

Spreadsheet cleanup and safe transformations (Excel-like work)

• AI can help restructure messy workbooks:
  • filtering/sorting
  • cleaning
  • removing duplicates
  • converting messy data into usable form
• Important: spreadsheet tasks differ by type:
  • changing appearance
  • changing records
  • changing formulas
• Suggested workflow:
  • decide duplication rules first
  • standardize names before deduping to avoid incorrect merges
  • ask for a cleaned copy + a short explanation of deletions
• Critical checks:
  • ensure dates/amounts sort correctly (display formatting ≠ actual values)
  • preserve formulas (don’t “hardcode” unless intended)
  • verify by modifying inputs and confirming dependent outputs change

Image/receipt extraction into structured tables

• ChatGPT Work can read photos (receipts and other documents) and convert them into sortable/usable tables.
• Best practice:
  • predefine required fields/columns (e.g., seller, date, total, currency, category)
  • preserve traceability by including source filename per extracted row
• Error handling / limits:
  • don’t force the AI to invent missing tax values
  • distinguish “blank/unknown” from “zero” meaningfully for finance/tax data
• Validation questions:
  • did each image create a line?
  • were any images duplicated or skipped?
  • were unrelated files excluded?

Research skepticism & evidence management

• Recommends prompting the model to maintain disagreement and not hide uncertainty.
• Emphasizes:
  • use current product documentation and relevant customer evidence
  • use correct market definitions
  • include publication dates and verify freshness (updated pages may still cite older info)
  • request a compact evidence table before a long report:
    • statement, source, date, uncertainty
• Large research strategy:
  • treat proposed segments/claims as hypotheses until validated
  • prefer fewer high-quality sources that directly support claims over huge link lists

Web browsing with verification

• For tasks needing data from websites or interactions, AI can use a browser:
  • local session browser vs cloud session browser
  • helps search/compare while keeping a verification trail (which pages it used)
• Credential guidance:
  • avoid entering sensitive info directly in conversation
  • use supported login flows when needed
• For outputs like travel plans:
  • ask for links and details
  • distinguish what’s conditional vs confirmed
• Narrow search parameters (dates, travelers, geography, constraints).

Integrations / connected apps (plugins)

• ChatGPT Work can connect to services like Google Docs, Slack, Gmail, calendars, etc.
• The video describes:
  • installing/connecting plugins
  • selecting/confirming accounts
  • clarifying ambiguity (e.g., which calendar among multiple calendars)
• Safety/accuracy checks:
  • confirm existing records first to avoid duplicates
  • distinguish actions like “add calendar event” vs “send invitations.”

Projects to manage long-running work

• Use Projects to keep shared sources and persistent instructions together across multiple tasks.
• Supports:
  • local projects (folders)
  • selecting which folder is the “main” context
• Good hygiene:
  • clean naming conventions (avoid “final final final”)
  • archive/organize completed work
  • avoid context ballooning (don’t overload folder context windows)

Computer history / access & privacy

• Notes a macOS computer history feature for supported accounts (Pro/Business/Enterprise):
  • can be enabled in settings
  • may require admin-managed access in managed environments
• Privacy cautions:
  • history processing isn’t guaranteed to remain purely on-device
  • enabling can expose accessible text (potentially sensitive content)
  • OpenAI recommends disabling if you may capture communications without others’ consent

Reusable “skills” via instructions

• Suggests creating skills: structured instruction templates for repeatable workflows.
• Skills are plain-markup instruction files (example: agents.markdown) intended for reuse.
• Guidance:
  • don’t convert every tip into a skill
  • start with a working method and generalize only what you truly repeat
  • if importing skills from others, verify safety and compliance (skills can include code/instructions)

Parallelization with multiple agents

• Advises using multiple agents when task parts can run independently (e.g., multi-store searches).
• Guidance includes:
  • give each agent a clear role/output format
  • have the “main task” reconcile conflicts into one unified recommendation
  • avoid parallelization when agents would wait on the same unresolved dependency or step on the same edits

Turning outputs into shareable websites/apps

• For interactive visualization or browsing collections, AI can create lightweight websites/apps rather than static text.
• Mentions:
  • separate preview vs private publish/deploy
  • ensure correct access controls
  • optionally schedule updates if content needs refreshing

Long-running tasks: “Goal” vs “Planning” + scheduling

• Goal
  • specify a result with clear completion conditions
  • can include status checks and stop conditions
  • supports pause/restore/suspend/continue patterns
• Planning
  • schedule repetitive runs (daily/weekly/etc.)
  • requires running once first, verifying quality, then scheduling the verified version
• Also mentioned:
  • event-triggered monitoring (Gmail/Slack/GitHub triggers)
  • time-based schedules
• Efficiency emphasis:
  • avoid overly frequent retries
  • add stopping conditions
  • only notify when action is needed

Sharing work safely

• Sharing strategies:
  • share the smallest useful artifact (file, site, or method)
  • understand permissions and limitations (e.g., Codex sharing creates a read-only snapshot, not a live shared branch)
  • snapshot access doesn’t automatically update with subsequent work
• Reminder: collaborators still need access to original sources/files to reproduce results.

Stopping/repairing runaway or incorrect tasks

• If a task is stuck or wrong:
  • check “status” / what completed / what needs approval
  • stop/pause/suspend the target rather than blindly retry
  • for connected actions, verify whether changes already occurred externally before repeating
• Safety note:
  • content read from web pages may be interpreted as instructions
  • ensure hostile/irrelevant text isn’t treated as a new request

---

Main speakers / sources (as stated in the subtitles)

• Nate B. Jones (main speaker; says “I’m Nate B. Jones”)
• OpenAI recommendations (referenced as an authority/source within the video)