The companion post covers what launched; this one covers what to run Monday · Sales / Marketing / Finance / Ops / Product / Engineering · Plan Mode · Scheduled Tasks · Usage optimization
Summary: On July 9, 2026, OpenAI shipped ChatGPT Work and folded Codex into the new ChatGPT desktop app. If you already know what it is, the next question is practical: what can I actually run at work tomorrow? OpenAI's own onboarding advice is straightforward—hand it a task you already know how to do. This guide follows that path: three principles that decide success or failure, Chat/Work/Codex routing, a five-step universal framework, copy-ready prompt templates for six roles, a Scheduled Tasks recipe library, seven usage optimization tactics, troubleshooting, and a 30-day ramp plan. For launch context and the Cowork comparison, read the companion post: ChatGPT Work Launch: Codex Merged Into the New Desktop App.
Before copying prompts, understand how ChatGPT Work differs from ordinary Chat:
| Principle | What it means | Practical tip |
|---|---|---|
| Describe outcomes, not steps | Work mode plans its own path; you define the finished product | ❌ "Open Salesforce, export data…" → ✅ "From @Salesforce, pull the last 30 days of opportunities and build a risk-annotated weekly report PPT" |
| Connect tools before assigning work | The plugin directory is Work's data layer | Authorize Gmail, Slack, Drive, and CRM before starting; use @app-name to specify sources explicitly |
| Plan Mode is your brake pedal | Complex jobs surface a plan first; you approve before execution | For high-stakes work—outbound email, financial reports, client deliverables—review every plan step before confirming |
The new ChatGPT desktop app runs three modes side by side. Using the wrong one wastes quota:
| Your need | Recommended mode | Why |
|---|---|---|
| Quick Q&A, brainstorming, single-turn copy | Chat | Lightweight and fast |
| Cross-app multi-step jobs, finished files, multi-hour runs | Work | Plugin integrations + Plan Mode + Computer Use |
| Code review, PR management, multi-repo development | Codex | Developer-specific workflow preserved |
| Weekly recurring, unattended background jobs | Work + Scheduled Tasks | Time- or trigger-based automation |
| Scenario | Recommended environment |
|---|---|
| Local file read/write, Computer Use, free-tier trial | Desktop (Mac / Windows) |
| Team collaboration, monitoring task progress on the go | Web / mobile (Plus and above) |
| Auto-generated sales meeting briefs + email notifications | Web Workspace Agent + scheduled runs |
| Local Excel reconciliation, batch folder processing | Desktop Work mode |
Treating Work like Chat—writing step-by-step instructions constrains the agent's planning ability and burns extra tokens.
Starting before plugins are authorized—the plan looks right, but execution cannot pull CRM or email data and falls back to guessing.
Skipping high-risk step review—outbound sends, file overwrites, and deletions are not caught in Plan Mode.
Windows users cannot run the macOS desktop natively—Computer Use, some plugin authorization flows, and Codex graphical validation require a real Mac environment.
Regardless of role, run every job through this sequence:
Connect plugins—authorize Gmail, Slack, Drive, CRM, and other tools from the plugin directory.
Define the goal and output format—use the prompt formula below to specify the deliverable (Docs / Excel / PPT / Sites).
Review Plan Mode—verify data sources, flag high-risk actions, and cut unnecessary steps.
Intervene mid-run—pause anytime to attach files or correct direction.
Accept the deliverable and iterate—polish the 80% draft to production quality, then convert it into a Scheduled Task.
[Role] + [Data source @plugin] + [Specific task] + [Output format] + [Constraints] + [Acceptance criteria] Skeleton example: You are a [role]. Pull [data type] from @Salesforce and @Gmail for [time range]. Complete [specific action] and output as [Google Docs / Excel / PPT / Sites]. Constraints: [do not modify source data / round amounts to two decimals / do not send external email]. When done, [notify me on Slack / save to a specific folder].
Confirm each item before execution:
The templates below draw on OpenAI's official examples, early tester feedback (Zapier, Nvidia, Virgin Atlantic, and others), and the Workspace Agent Cookbook. Swap @plugin-name to match your stack.
Scenario A: Auto-generated client meeting briefs (daily schedule)
Pain point: reps spend 1–2 hours daily assembling client background. Work approach: scan the calendar → pull CRM notes → search news → generate briefs.
Create a scheduled task that runs every weekday at 4:00 PM. 1. Check my @Google Calendar for client meetings tomorrow (exclude internal meetings) 2. For each client meeting: - Pull account notes and interaction history from the last 30 days via @SharePoint / @Salesforce - Search for public news and executive updates about the company from the last 30 days - Write a 2–3 sentence background summary for each external attendee 3. Generate a 2–3 page brief for each meeting and save as @Google Drive documents 4. Send me a summary @Gmail with links to each brief Output format: email subject "Tomorrow's Client Meeting Briefs — [date]", body as a table (Client | Meeting time | Key topics | Brief link)
OpenAI internal case: a sales team turned a single discovery conversation into a customized PoC proposal within 24 hours—a process that traditionally took weeks.
Scenario B: Account command center (Sites + daily refresh)
Based on all opportunities, contacts, and recent activity for [account name] in @Salesforce: 1. Create an interactive account command center (Sites) containing: - Pipeline overview (stage, amount, expected close date) - Key signals from the last 7 days (email, meetings, support tickets) - Recommended next actions (priority ranked) 2. Set a Scheduled Task to refresh the Site every weekday at 8:00 AM 3. DM me on @Slack when there are significant changes Constraints: do not auto-send any external email; amounts must match CRM source data.
Scenario C: Lead review and pipeline repair (adapted from Zapier case)
Analyze new leads from the last 30 days in @Salesforce and their follow-up history, cross-referenced with sales correspondence in @Gmail. Find: 1. Leads with no follow-up after 48 hours (grouped by source) 2. Break points in the follow-up chain (where response rate drops sharply) 3. Estimated pipeline loss in dollar terms Output: - Excel detail sheet (Lead ID | Source | Last follow-up | Break type | Recommended action) - 1-page executive summary PPT highlighting seven-figure opportunity risk - A weekly review process suitable for conversion into a Scheduled Task
Scenario A: Research → Brief → multi-market assets (end-to-end pipeline)
I have uploaded the following customer research: [attachment / @Google Drive link] Complete an end-to-end marketing workflow: Phase 1 — Brief: - Extract target audience, core pain points, and competitive positioning - Output a Campaign Brief (Google Docs) with message pillars and channel recommendations Phase 2 — Asset generation: - From the Brief, generate: 1 acquisition email, 3 LinkedIn posts, 1 landing page copy outline - Save to @Google Drive folder "Campaign / [product name]" Phase 3 — Regional adaptation: - Adapt core assets for US, Europe, and APAC (language, cultural references, compliance wording) - Flag sensitive phrasing that requires human review in each version Pause after each phase and wait for my confirmation before proceeding.
Scenario B: Slack / Teams activity synced to meeting agenda (Scheduled Task)
Set a scheduled task to run every Monday at 7:00 AM: 1. Summarize important discussions from the last 7 days in @Slack #product-launch and @Microsoft Teams "Go-to-Market" channel 2. Extract: decisions made, open questions, blockers that need alignment in the meeting 3. Update the "Weekly Meeting Agenda" document in @Google Drive (preserve version history) 4. Post a summary of 5 items or fewer in @Slack #leadership Constraints: cite only publicly discussed content; do not leak messages marked confidential.
Scenario A: Month-end variance analysis (OpenAI internal validated scenario)
Internal result: month-end close and forecasting compressed from days to hours.
Help complete [month] budget variance analysis: 1. Pull corresponding tables from @Google Drive "Finance / Actuals" and "Finance / Forecast" 2. Create a reconciliation workbook in @Google Sheets: - Summarize actual vs forecast variance by department - Flag line items with variance >5% or >$50K - Preserve all original formulas; do not overwrite source files 3. Draft a performance narrative (Google Docs) explaining likely causes by Revenue / COGS / OpEx 4. Build a 5–8 page management report PPT (with charts, following the attached template style) 5. List 3 key judgment calls that require manual finance sign-off Constraints: do not modify any source data; cite source cell for every figure.
Scenario B: Invoice and payment reconciliation (first gate for AP automation)
You are an accounts payable specialist. Compare the following two datasets: - Payment register: [@Google Drive link] - Invoice list: [@Google Drive link] Flag these anomalies (return as a table): | Issue type | Vendor | Invoice # | Amount | Recommended action | - Amount variance >2% - Missing tax ID - Duplicate invoice number - Vendor name mismatch Do not initiate payments; output a review sheet for manual verification only.
Scenario A: Daily dashboard change monitoring (Scheduled Task)
Run automatically every weekday at 6:30 AM: 1. Access [internal dashboard URL / @SharePoint report page] 2. Compare against yesterday's snapshot; extract significant changes (>10% swing or new red indicators) 3. Generate a 1-page morning brief (Google Docs) structured as: - TOP 3 items to watch today - Metric change table - Recommended follow-up owners 4. Send via @Gmail to ops-leads@company.com If the dashboard is unreachable, tell me during Plan Mode—do not fabricate data.
Scenario B: Customer feedback clustering → product prioritization
Monitor new customer feedback from the last 14 days across: - @Slack #customer-feedback - @Gmail label "NPS-Detractor" - @Google Drive "Support Tickets Export" 1. Cluster feedback into 5–8 themes (with representative quotes) 2. Score priority by frequency × impact × implementation difficulty 3. Output a product evaluation backlog (Notion / Google Docs format) 4. Set a Scheduled Task to refresh the document every Friday Constraints: anonymize feedback; do not include customer names.
Scenario A: Cross Jira + GTM launch readiness review (adapted from Nvidia case)
Run a launch readiness review for [product/feature name]: 1. Pull associated Epic / Story completion status and open blockers from @Jira 2. Pull the corresponding GTM plan from @Google Drive "GTM Plans" and check key milestones 3. Extract unresolved discussions from the last 7 days in @Slack #product-launch 4. Output a Launch Readiness report (Google Docs): - Readiness score (red / yellow / green) - Blocker list (owner | due date | risk level) - Recommended Go / No-Go judgment with rationale Do not auto-update Jira status; flag high-risk items for human decision.
For engineering, use Codex for code and Work for cross-team documentation. Both live in the same desktop app—switch modes without changing tools.
Scenario A: PR review + release notes (Codex-led)
In Codex mode: 1. Review PR #123 in [repo/name], focusing on [security / performance / test coverage] 2. Leave line-by-line review comments in the PR sidebar 3. If approved, generate a Release Notes draft Then switch to Work mode: 4. Format Release Notes as a @Confluence page 5. Draft an announcement for @Slack #engineering (do not auto-send)
Scenario B: Multi-repo issue summary weekly report (Codex multi-repo capability)
In Codex mode, across [frontend-repo] and [backend-repo]: 1. Summarize merged PRs this week and open P0/P1 issues 2. Generate an engineering weekly report in Markdown Switch to Work mode: 3. Convert to Google Docs and insert this week's burndown chart (pulled from @Jira) 4. Set a Scheduled Task to auto-generate every Friday at 5:00 PM
Four high-frequency scheduled task patterns OpenAI recommends—you can adapt them directly:
| Recipe | Trigger | Task description | Best for |
|---|---|---|---|
| Monday agenda refresh | Every Monday 07:00 | Summarize Slack activity → update agenda doc | Marketing / Ops |
| Daily metrics morning brief | Every weekday 06:30 | Visit dashboard → compare yesterday → email summary | Ops / Finance |
| Feedback clustering weekly | Every Friday 16:00 | Multi-channel feedback → theme clusters → priority list | Product |
| Account activity daily | Every weekday 08:00 | CRM changes → refresh Sites command center | Sales |
Set up a Scheduled Task: - Frequency: [daily / every Monday / 1st of month / when keyword appears in @Slack channel] - Time: [timezone + specific time] - Action: [specific workflow description] - Notification: [Slack channel / email / none] - Manual approval: [which steps require my sign-off first]
ChatGPT Work and Codex share a single usage pool (not a flat monthly feature fee). The same workflow, designed differently, can cost 5× more or less.
| Factor | Impact on usage |
|---|---|
| Number of steps | More steps, higher consumption |
| Context size | More documents and emails pulled, higher consumption |
| Output length | Output tokens cost roughly 6× input tokens |
| Cache hits | Re-reading the same document costs about 1/10 of fresh input |
| Model selection | GPT-5.6 complex reasoning costs more than lightweight tasks need |
Draft in Chat mode first, then hand a trimmed version to Work for execution
Cut redundant steps in Plan Mode, especially repeated pulls from the same data source
Reuse the same template document in Scheduled Tasks to benefit from cache discounts
Keep output concise: "table + 3 bullet summary" beats "full narrative report"
Split large jobs: Phase 1 confirms direction → Phase 2 generates the deliverable, avoiding expensive full reruns
Free tier users: run small desktop tasks first to measure consumption before scaling
Enterprise teams: set workspace / group / per-user quotas in the Admin Console
1. Pick a real task you already know the manual time for (e.g., month-end variance table, usually 2 hours by hand) 2. Run it once in Work mode with Plan Mode enabled; record step count 3. After execution, check consumption against your plan's included usage 4. Estimate: if run daily / weekly, is monthly consumption within budget? 5. If high → optimize per Section 5.2 and rerun to compare
| Problem | Cause | Fix |
|---|---|---|
| Work mode cannot find installed Codex projects | App migration not completed | Update the Codex app → it becomes the ChatGPT desktop client; if broken, reinstall from chatgpt.com/download |
| Plugin authorized but data still missing | Insufficient scope or wrong @app-name spelling | Check authorization scope in the plugin directory; write @Salesforce explicitly, not a generic "CRM" |
| Plan looks right, execution goes off track | Stale context files or AI inference | Pause and correct mid-run; provide key data via attachments or explicit links |
| Scheduled task did not trigger | Computer asleep / desktop not logged in | For long-cycle jobs, use web Workspace Agent; desktop Scheduled Tasks need the device awake |
| Usage higher than expected | Long output, repeated pulls, too many steps | Apply Section 5 optimizations; Enterprise admins can set limits in Admin Console |
| Unclear whether to use Work or Cowork | Different workflow types | Cloud SaaS collaboration → Work; local folder batch processing → Cowork (see companion post comparison) |
| Phase | Goal | Action |
|---|---|---|
| Week 1 | Master a single task | Pick one familiar job; run it manually in desktop Work mode 3 times; practice Plan Mode review |
| Week 2 | Deep plugin integration | Connect 3 core tools (email + collaboration + files); complete one cross-app end-to-end delivery |
| Week 3 | Automation | Convert the Week 1 task into a Scheduled Task; verify 3 successful triggers |
| Week 4 | Team rollout | Build a role-specific prompt template library; Enterprise teams sync quota settings with admins |
Pick a task you already know well enough to judge output quality. OpenAI recommends month-end variance analysis, marketing briefs, and sales meeting prep—you can verify results quickly.
Focus on "data source + output format + constraints"—usually 150–400 words is enough. Do not write every manual step; that is what Work mode automates.
Desktop Scheduled Tasks require your device to be online. For true unattended background runs, use the web Workspace Agent on Plus and above.
Work mode is the personal agent you use directly inside ChatGPT. Workspace Agent is a team-built, shared, centrally managed automation agent inside Business or Enterprise—with Admin Console governance. The underlying technology is similar; the entry point differs.
Treat them as "80% drafts." Always manually verify financial figures, customer names, and external-facing statements before use.
Desktop Work mode is available on the free tier with usage limits. Start with lightweight tasks like "Scenario B: Invoice Reconciliation" before scaling to long-running automation.
ChatGPT Work earns its keep not by existing on a feature list, but by taking over workflows you are already tired of doing manually. The fastest ROI is not reading more launch posts—it is picking one task you know cold, running it three times by hand, tuning the prompt, then automating it.
If your primary machine is Windows or Linux but you need the macOS desktop app, Computer Use, or Codex graphical validation, buying a Mac is expensive and local VMs often stall on permission dialogs. Rent a VNCMac remote Mac by the hour: open a graphical macOS session over VNC, install the ChatGPT desktop app, connect plugins, and run any role template from this guide—then stop renting when the project ends. See Mac Mini M4 plans to get started.