ChatGPT Work: Summary & Action Guide

Source & Credit

Summary and action checklist distilled from this walkthrough video. Original post: x.com/i/status/2075344215270101080

Deeper platform + power-user workflows (Work vs Codex, computer use, loops, Sites, multi-thread): ChatGPT + Codex Super App (part 2).

What Changed

ChatGPT shipped a major update with three pieces that matter for day-to-day work:

1. ChatGPT Work — an agent mode that plans in the background, gathers context, and hands you finished work (files, decks, spreadsheets, hosted sites) instead of another chat reply.
2. Three models tuned for Work: Sol (most capable), Terra (balanced default), Luna (fastest / cheapest).
3. Desktop app upgrades — same Work + model stack, plus folder access so the agent can read/write local files (web Work downloads files or gives shareable hosted links instead).

Closest competitor pattern: Claude Cowork (and open alternatives like Eigent / OpenWork). Same idea — move from “talk about the work” to “do the work.”

What Work Is Good At

Models: Sol, Terra, Luna

These ship with the Work update (around the ChatGPT 5.6 generation) and are the ones you pick for agent tasks. Paid plans unlock model choice and Work; free / Go typically stay on Terra.

Effort settings matter as much as the model. Sol + max reasoning + fastest speed burns credits hard and can take 15–30+ minutes on big jobs. Prefer:

• Hardest tasks: Sol + max (or extra-high) only when quality is worth the wait.
• Most Work tasks: Terra + high effort — everyday default.
• Sol day-to-day: medium or high effort, not max every time.

Pro tip from early access: before a big run, start a normal ChatGPT chat, paste the prompt + describe the files, and ask which model + effort it recommends. Use that instead of always maxing Sol.

How to Access Work

1. Open ChatGPT on the web or the Mac/Windows desktop app (paid plan for Work + Sol).
2. Switch from the regular Chat tab to the Work tab — finished deliverables, not just conversation.
3. Pick Sol / Terra / Luna and reasoning effort before you send.
4. Optional: attach a project (web) or local folder (desktop) so outputs land in one place.
5. Optional: enable plugins / connectors (Gmail, calendar, Slack, company context).
6. Watch the outputs panel for files, hosted sites, and sources (your uploads/inputs).

Web vs desktop: both can run Work. Desktop can open folders on your machine and write files there. Web returns download links and shareable hosted site URLs.

Action: Five Workflows to Run

Copy the intent below into Work. Adjust files and brand context to your stack. Times are from the source demo — your runs will vary with effort and queue.

1. Data audit → hosted dashboard

Model used in demo: Sol (recommended by Chat after describing the task).

Action prompt pattern:

• Analyze every file I upload (or every file in this folder on desktop).
• Validate data, clean it, run an audit.
• Publish a hosted dashboard/website with the cleaned data (e.g. by region) and useful filters.

What you get: hosted site (shareable .site link), not just a canvas mock — typically ~15 minutes in the demo. First draft is good enough to share; iterate with a follow-up Work chat if needed.

2. Trip / personal hub (mobile-first hosted site)

Model used in demo: Terra + medium effort (Chat recommended Terra over Sol).

Action prompt pattern:

• Upload itinerary notes, bookings, budgets, constraints (4+ source files in the demo).
• Build a mobile-first hosted trip hub with day-by-day itinerary, budget, and shareable link for family/team.

Nice UX detail: Work can pause and offer 2–3 design directions — pick one before it finishes the site. Check mobile layout by narrowing the browser; that was a hard requirement in the prompt.

3. Style-matched PowerPoint from data

Model used in demo: Sol + max reasoning.

Action prompt pattern:

• Upload a sample deck (style reference) + a CSV / spreadsheet (content).
• Ask for a 10–12 slide presentation that mirrors the sample’s design system and includes all key metrics/story from the data.

What you get: downloadable PowerPoint that visually matches the template. Demo produced 11 slides in the requested range.

4. Full marketing launch kit (one agent run)

Model used in demo: Sol + extra-high effort (~33 minutes). Max was recommended; extra-high saved credits for the recording.

Action prompt pattern (single message, multi-deliverable):

• Sources: brand guidelines, product brief, customer feedback.
• Outputs in one go: launch strategy (e.g. 30-day plan), 10-slide deck, 5-email sequence, 12 social posts, 3 ad concepts, plus a hosted marketing site.

Why it’s impressive: step-by-step agent execution across assets that stay on-brand together — closer to a campaign team than a single chat reply. Same multi-asset idea shows up in AI content marketer / Content OS workflows, just inside ChatGPT Work instead of a custom stack.

5. Daily command center from inbox + calendar + Slack

Model used in demo: Luna + max (for speed in the recording). Real use: Terra or Sol + connectors.

Action prompt pattern:

• Review inbox, calendar, Slack (via plugins — or temporary exports if demoing without live access).
• Produce a one-page command center: top priorities, urgent items, decisions needed today.
• Export as PDF; then follow the agent’s suggestion to connect live Gmail/Slack and re-run in a fresh Work chat.

Best practice: connectors > file dumps. Live context makes the one-pager actually useful every morning.

Credit Hygiene (Don’t Run Out Mid-Day)

• Do not default Sol + max for every task — paid plans still hit rate/credit walls.
• Ask Chat once for a model recommendation on novel, large jobs.
• Use Terra + high for routine Work; reserve Sol max for decks, multi-asset launches, hard data + product logic.
• Long high-effort runs can take 15–30+ minutes — start them when you can leave the tab open.
• Desktop folder projects help you reuse the same workspace without re-uploading.

Quick Start Checklist

1. Confirm a paid ChatGPT plan if you need Work + Sol.
2. Open the Work tab (web or desktop).
3. For the first week: run workflow #5 (command center) daily and workflow #1 or #2 once for a shareable site.
4. For client/brand work: keep a sample PPT + brand PDF ready for style matching (#3 and #4).
5. Compare the same prompt against Claude Cowork / Eigent if you care about cost, accuracy, and credit burn side-by-side.
6. Level up prompting context with Andrew Ng’s 2-hour prompting course — Work multiplies good prompts; it does not replace them.

Related Guides

Content summarized for action from the original walkthrough. View the original video post on X →