The ChatGPT Work Guide

What ChatGPT Work actually is

On 9 July 2026, OpenAI added a third mode to ChatGPT — alongside Chat there's now Work. Instead of prompting it line by line, you hand it an outcome. It writes a plan, you approve it, then it goes away and does the job — pulling from your connected apps, checking in when it needs you, and working for minutes or hours. What comes back isn't a chat reply. It's the finished file.

💬 Regular ChatGPT

You prompt, it replies
You copy-paste into your real tools
You assemble the final thing yourself
Brilliant for thinking, drafting, editing

🛠️ ChatGPT Work

You describe the outcome
It plans — you approve — it builds
Hands back the finished deck, sheet or doc
Built for reports, proposals and data jobs

🗺️

Plan mode

It proposes a step-by-step plan before touching anything. Approve it, edit it or redirect it. This is your control panel — keep it on.

📦

Finished files

Documents, spreadsheets, slide decks, dashboards — real files you can send, not chat text you have to reformat. It can even publish simple hosted websites (Sites, still in beta).

🔌

Connected apps

Wire in Gmail, Google Drive, Calendar and more so it works from your real information instead of guessing. This is the step most people skip — the full walkthrough is next.

🔁

Scheduled runs

Set a brief to re-run monthly or every Monday morning. The competitor scan and the briefing above are built for exactly this.

Already using Agent mode, Deep Research or Tasks?

Work is the step up from all three. Deep Research hands you a report — Work builds the deck from it. Tasks send you a scheduled note — Work's scheduled runs do the whole job. Agent mode babysits one browsing session — Work runs multi-step jobs for hours.

What you need

A paid plan — Plus (US$20/month) is the floor. There's no extra subscription for Work itself, but usage is metered against your plan rather than unlimited (more on that in the traps section). It reached Pro and Enterprise accounts first, with Plus and Business following in the days after launch — if you can't see it, update your app and check again.

Connect your tools — the step everyone skips

Here's the thing nobody tells you: Work mode is only as good as what it can reach. Unconnected, it's guessing — polished, confident guessing. Connected to your actual Gmail, Drive and Calendar, it works from your real numbers, your real enquiries, your real week. Same brief, completely different result.

❌ Basic prompting

"Write my monthly client report"
It invents plausible-looking numbers
Generic wins, filler slides
You spend an hour fixing it — 2023 all over again

✅ A connected brief

"Build [Client]'s report deck from @Google Drive"
It reads your actual spreadsheet
Your deck order, your talking points
You review a finished file instead of rebuilding it

Where connections live

Profile icon → Settings → Apps (older versions call it Connectors). Click Connect next to an app, sign in on that app's own login page, and choose what it's allowed to see. That's it — the app now shows up in the ⊕ tools menu, and you can @-mention it in any brief.

Start with these three. They power every brief in this guide, and connecting them takes about ten minutes total:

📧

Gmail

Lead triage, enquiry search and drafted replies from your real inbox. It drafts freely — but every send needs your individual approval, so nothing goes out without you clicking yes. Powers briefs 2 and 6.

📁

Google Drive

Reports and decks built from your actual sheets and docs. When you connect, choose specific folders — it doesn't need your whole Drive, just the client work. Powers briefs 1, 3 and 5.

📅

Google Calendar

Weekly briefings and call-prep notes from your real schedule. It's read-only — it can read your week, it can't move a single meeting. That's a feature, not a gap. Powers brief 6.

🪟

On Microsoft?

Same steps, different names — connect Outlook instead of Gmail, OneDrive instead of Drive. The briefs work identically; just swap the @-mentions.

Then @-mention them in your briefs

Type @ in any brief and pick the app — @Gmail, @Google Drive, @Google Calendar — exactly like tagging someone on Instagram. That one character is the difference between "go and look at my actual inbox" and "make something up". It searches one connected app at a time, so if a job needs your inbox AND your calendar, name both. Every brief in the next section has the @-mentions already written in.

What connections can't do (yet)

Three honest caveats before you plan your empire around them. Gmail drafts anything, but each send needs your one-click approval — for client-facing email, drafts-only is the setting you want anyway. Calendar is read-only — brilliant for briefings, useless for rescheduling. And it searches one app per look-up, so a vague "check my stuff" brief goes nowhere — name the apps. Design your briefs around those three and you won't hit a wall.

Notion, Slack, HubSpot, Canva…?

There's a directory of 1,400-odd apps in there. Connect one when a job actually needs it — not before. Every connection is access you have to remember you granted, and you can disconnect anything, any time, in Settings → Apps. While you're in Settings: on personal plans, go to Data controls and switch off the "improve the model" toggle before you connect a single thing.

Set it up in 10 minutes

1

Update, then find the Work tab

OpenAI shipped a new desktop app with the launch (Chat, Work and Codex in one). Old versions hide the feature entirely — update first, or just use chatgpt.com in your browser.

2

Connect your starter three

Gmail, Google Drive, Google Calendar — the walkthrough is in part 2 above. It's tempting to connect everything in the directory — don't. Every connection is access you'll forget you granted. Add more later when a job needs it.

3

Leave Plan mode and approvals ON

In week one there were reports of the flagship model deleting files and swapping resources without asking. Approvals are your seatbelt — the small pause is the feature, not the friction.

4

Pick the cheaper model for routine jobs

Work runs on three models — Sol, Terra and Luna. Sol is the heavy thinker and burns roughly five times the credits of Luna. Terra or Luna handle the everyday briefs just fine; save Sol for genuinely hard problems.

5

Run ONE small job first

Start with the spreadsheet brief. Watch how it plans, how it checks in, what it costs against your limits — then decide what earns a regular spot.

A paid ChatGPT plan — Plus (US$20/month) is enough

The updated app, or chatgpt.com in a browser

10 minutes and one real job to hand it

Enterprise pricing — the floor is a Plus plan

Tech skills — if you can write a text message, you can write a brief

A separate subscription — Work is included, metered by usage

The 6 outcome briefs

These aren't prompts — they're briefs. Work mode wants three things: the finished outcome you're after, the context it needs, and your rules. The context is where your connected tools earn their keep — every brief below names the apps it pulls from, with the @-mentions already written in. Fill the [BRACKETS], paste, and approve the plan it shows you before it runs.

Read the chips

Each brief shows the tools it uses. 🔌 means it pulls from a connected app (that's the part 2 walkthrough). 📎 means you just attach a file — no connection needed. Haven't connected anything yet? Start with a 📎 brief and wire the tools in after.

1

The monthly client report deck

The highest-value one. If you send clients a monthly report, this turns a two-hour job into review-and-send.

🔌 @Google Drive📎 or attach the export

Use Plan mode: show me your step-by-step plan and wait for my approval before doing anything. OUTCOME: My monthly client report for [CLIENT NAME], covering [MONTH], as a finished slide deck. CONTEXT: - Numbers: @Google Drive → [name the exact sheet, e.g. "March results — Client X"] (or attach the export) - Wins and talking points: [attach or paste your call notes] - Deck order: results summary → what we did → what's working → what's next → recommendations RULES: - Use only the numbers in my file. If something's missing, ask me — never estimate. - Plain language. No jargon, no filler slides. - Anything unusual in the data gets its own slide, flagged clearly. - Show me the draft before you finalise anything.

2

Lead triage with drafted replies

Needs your email connected — that's the part 2 walkthrough. Then this sorts a week of enquiries in one run.

🔌 @Gmail (or @Outlook)

Use Plan mode: show me your plan and wait for my approval. OUTCOME: Every enquiry from the last 7 days triaged, scored and answered — as drafts for my review. Do not send anything. CONTEXT: - Check @Gmail (or @Outlook) for new enquiries only — ignore everything else. - My ideal client: [describe them: industry, size, what they want, budget signals] - Red flags: [e.g. asking for free work, budget under $X, outside my niche] - How I write: [paste 2-3 sentences from a real reply so it can match your voice] OUTPUT: A table — name, what they want, score out of 10, why — then a drafted reply for each lead. RULES: Drafts only. Never send, delete or file anything.

3

Discovery call → proposal pack

One transcript in, a proposal and kickoff checklist out. The rules stop it inventing services you don't sell.

📎 attach the transcript🔌 @Google Drive if it lives there

Use Plan mode: show me your plan and wait for my approval. OUTCOME: Two finished documents from the attached call transcript: (1) a proposal, (2) a kickoff checklist. CONTEXT: - Transcript: [attach it, or @Google Drive → the file name] - My services and prices: [list them exactly] - Proposal structure: their problem in their own words → what I'll do → timeline → investment → next step RULES: - Quote their actual words in the problem section. - Only offer services from my list, at my prices. Never invent an offer or a discount. - Anything unclear from the call goes in an "open questions" list at the end — don't guess.

4

The competitor scan that runs itself

Schedule this one. It re-runs monthly and takes five minutes to read instead of an afternoon to research.

🌐 built-in browser — no connection needed🔁 scheduled monthly

OUTCOME: A competitor briefing document I can read in 5 minutes. Schedule this to re-run on the 1st of every month. CONTEXT: - My business, one line: [what you do and for whom] - Competitors: [names + websites + key social profiles] EACH RUN: Check their sites, pricing pages and socials for changes since last time — new offers, price moves, new content angles, anything they're pushing hard. OUTPUT: One document: what changed, what it means for me, and ONE suggested response. If nothing changed, say so in one line — don't pad it out.

5

Spreadsheet clean-up + mini dashboard

The safest first run — nothing client-facing at stake, and you'll see how Work plans, checks in, and ships a file.

📎 attach the export🔌 @Google Drive works too

Use Plan mode: show me your plan and wait for my approval. OUTCOME: This spreadsheet cleaned, plus a simple dashboard tab. CONTEXT: [attach the export — sales, bookings, subscribers, whatever you've got] CLEANING: Remove duplicates, fix name casing, standardise dates to DD/MM/YYYY, and split anything jammed into one column. DASHBOARD TAB: Totals by month, top 5 [products / clients / sources], and anything unusual flagged in its own list. RULES: - Show me 5 sample rows of cleaned data BEFORE you build the dashboard. - Never delete rows — move anything dodgy to a "check these" tab so I can decide.

6

The Monday morning briefing

Scheduled, read-only, under 300 words. Your week, sorted before your coffee goes cold. This is the one that uses both connections at once.

🔌 @Gmail🔌 @Google Calendar🔁 scheduled weekly

OUTCOME: A Monday-morning briefing I can read with my first coffee. Schedule this for every Monday at 7am. CONTEXT: Check @Gmail for anything I haven't answered from last week, and @Google Calendar for this week's calls and deadlines. OUTPUT, one page max: 1. The 3 things that need me first (and why) 2. This week's calls, each with a one-line prep note 3. Anything I've left hanging that's about to become a problem RULES: Under 300 words. Blunt beats polite. Read-only — don't touch the emails or the calendar.

Fill the brackets or expect mush

The brackets ARE the brief. Skip them and you'll get the same generic output that made you shrug at ChatGPT in 2023. And if a brief names @Gmail or @Google Drive, connect it first — an @-mention with nothing behind it is just decoration. Two minutes of context is the difference between "meh" and "send it".

Plugins, skills and automations

Once you've run a few briefs, three features turn Work from handy into "this runs part of my week". They're the level-up, so leave them until your starter three are connected and a brief or two is under your belt.

🧩

Plugins

The app directory is now called Plugins. Each one bundles a connection plus ready-made skills and task templates for that tool, so installing Slack gets you channel summaries out of the box, not just access.

📐

Skills

A skill is your SOP written once so Work follows it every time. Same steps, same format, no re-explaining. Think "how I build a client report" saved as a recipe it reaches for by name.

⏰

Automations

Turn any brief into a job that runs on a schedule, or when something changes. Briefs 4 and 6 are built for this. Up to 10 running at once on a paid plan.

Finding plugins

In Work mode, open Plugins, search or browse, hit install, sign in if it asks, then start a fresh chat. Worth a look for a small business: Slack, Canva, Notion, Dropbox and Zoom. Same rule as your connections though, add one when a job needs it, not for the collection. Plugins only live in Work mode, not the Chat tab or the phone app.

Building a skill is the same as writing a good brief, you just save it so you never write it again. The quickest way in is to ask Work to build it for you:

Build me a skill called "[NAME, e.g. Monthly client report]". WHEN TO USE IT: [e.g. any time I say "build the report for [client]"] STEPS: 1. Pull the numbers from @Google Drive → [the sheet] 2. Follow my deck order: results → what we did → what's working → what's next 3. Flag anything unusual on its own slide FORMAT: A finished slide deck, plain language, no filler. RULES: Only my real numbers. If something's missing, ask me, never estimate.

Making an automation

Say what should happen and when. My demo was "create an automation for new membership gifts", it watches for a new member and drafts the welcome each time. Paste a brief you already like, then add: "Turn this into an automation. Run it [every Monday 7am / when a new lead lands]. Draft only, never send without me."

The honest bit on skills

At launch, saved Skills are a workspace feature, so Business, Enterprise and Education plans see them first. On a personal Plus plan you might not have the Skills library yet. No drama, keep your favourite briefs in a note or a Project and paste them in. You get the same result while it rolls out.

The traps that burn your credits

Launch week was messy. Demand doubled in 48 hours, early users torched their usage limits in a handful of prompts, and by 12 July OpenAI publicly admitted "we didn't get everything quite right" and loosened the caps. Temporarily. Here's what the first wave learned so you don't have to.

Trap 1 — the metered pool is real

Work isn't unlimited. It draws from a shared usage pool, and heavy jobs weigh more. The current loose limits are a launch-week apology, not the plan — treat credits as finite from day one and start with small jobs.

Trap 2 — Sol eats roughly 5x the credits

The flagship model is brilliant and expensive. If you leave it as the default for every routine brief, your limits vanish. Terra or Luna for everyday work; Sol when the job is genuinely hard.

Trap 3 — polished isn't the same as right

A finished-looking deck is the most dangerous kind of error, because you stop checking. One wrong source early poisons every slide after it. Spot-check the numbers and sources before anything goes to a client — Deloitte had to refund part of a $440k government contract over AI-fabricated citations. You don't get to make that mistake once.

Trap 4 — connector creep

It's very easy to grant broader Gmail and Drive access than you meant to. Stick to the starter three from part 2, audit monthly, and disconnect anything a job no longer needs. If you handle client-confidential information, the Business plan (which doesn't train on your data by default) is the clean upgrade.

Trap 5 — don't publish client data to Sites

The Sites beta makes putting a page on the actual internet one approval away. Lovely for a portfolio page; a very bad place for a client dashboard until you've thought about who can see it.

What it won't do (yet)

The honest list

Run your business unattended

It pauses for approvals, logins and payments, and usage is capped. It's a supervised assistant, not an employee.

Make your decisions

It will happily write the pricing page so you can avoid deciding what you're worth. That's dodging the decision, not speeding it up.

Tax, legal, compliance

Confident, plausible, unaccountable. "The AI told me" is not a defence anyone accepts.

Deep recurring automation

Multi-system plumbing that runs itself — order flows, sync jobs — still belongs in proper automation tools, not an agent you supervise.

One more truth: research on 1.5 million real conversations found most workplace AI use is still writing and thinking-partner work — that's where the reliable wins are. Work mode adds the shipping. It doesn't replace the thinking.

And remember this is week-two software. Limits are in flux, the desktop app is mid-redesign, and features will move around. The owners who win aren't the ones who master today's buttons — they're the ones with their tools connected and briefs ready to paste when it settles. You now have both.