2026 ChatGPT Work: 6 ролей, prompt-шаблоны и автоматизация Scheduled Tasks
После релиза ChatGPT Work (9 июля 2026) инженерный вопрос звучит иначе: какой pipeline завтра отдать agent? OpenAI рекомендует стартовать с задачи, которую вы уже умеете верифицировать. Ниже — шесть ролей (sales, marketing, finance, ops, product, engineering) с copy-paste prompt, Plan Mode checklist, Scheduled Tasks recipes и consumption tuning. Архитектура продукта и Metal-throughput — в сестринской статье о релизе.
1. Три anti-pattern: Chat ≠ Work
- Wrong mode → 2× consumption. Multi-step cross-app в Chat или trivial Q&A в Work сжигает shared billing pool. Day-one cost trap — плохой routing Chat/Work/Codex.
- Prompt как runbook. Work планирует path сам. «Open Salesforce, export…» фиксирует graph и увеличивает retry count.
- Unattended cron без safety gate. Auto-send email, overwrite shared doc, wrong month pull — high-risk ops должны быть explicit constraints в Plan Mode и Scheduled Task config.
2. Три принципа до первого run
| Принцип | Механизм | Ops hint |
|---|---|---|
| Outcome, не steps | Agent строит execution graph | ❌ «Open CRM…» → ✅ «@Salesforce 30d pipeline → risk-tagged weekly PPT» |
| Tools first | Plugin catalog = data plane | OAuth Gmail/Slack/Drive; @app explicit в prompt |
| Plan Mode = circuit breaker | Plan → approve → execute | External comms, finance, customer deliverables — line-by-line approve |
2.1 Mode routing: Chat / Work / Codex
| Workload | Mode | Rationale |
|---|---|---|
| Q&A, brainstorm, one-shot copy | Chat | Low latency, min steps |
| Cross-app, artifacts, multi-hour jobs | Work | Plugins + Plan + Computer Use |
| PR review, multi-repo dev | Codex | Dev-native graph |
| Recurring unattended | Work + Scheduled | Cron/trigger automation |
2.2 Desktop vs Web execution surface
| Scenario | Surface |
|---|---|
| Local FS, Computer Use, Free trial | Desktop Mac/Win |
| Team visibility, progress polling | Web/Mobile Plus+ |
| Sales brief cron + email notify | Web Workspace Agent |
| Local Excel recon, folder batch | Desktop Work |
3. Five-step pipeline и prompt formula
1. Connect plugins → 2. Define outcome/format → 3. Approve Plan Mode → 4. Mid-flight correction → 5. Accept & iterate3.1 Work prompt formula
[role] + [@plugin data source] + [task] + [output format] + [constraints] + [acceptance criteria]Skeleton: As [role], pull [period] [datatype] from @Salesforce and @Gmail. Produce [action] as [Docs/Excel/PPT/Sites]. Constraints: no source mutation / 2-decimal amounts / no external send. Notify via Slack or save to path X.
3.2 Plan Mode approval checklist
- Correct data source (wrong customer/month)?
- High-risk: send/delete/overwrite?
- Output matches team template?
- Steps removable to cut consumption?
- Human-in-the-loop required?
4. Sales: три scenario prompt
Based on OpenAI refs, Zapier/Nvidia/Virgin Atlantic early feedback. Swap @plugin под ваш stack.
4.1 A: Meeting brief cron (daily)
Pain: 1–2 h/day manual account research. Work: calendar → CRM → news → brief artifact. OpenAI internal: Discovery → custom PoC in 24h.
Scheduled Task: weekdays 16:00.
1. Tomorrow's @Google Calendar customer meetings (exclude internal)
2. Per meeting:
- @SharePoint/@Salesforce: 30d account notes & touches
- Public news & exec moves (30d)
- 2–3 sentence background per external attendee
3. 2–3 page brief per meeting → @Google Drive doc
4. @Gmail summary with links
Subject: "Tomorrow customer brief — [date]"; body table (account | time | topics | link)4.2 B: Account command center (Sites + daily delta)
From @Salesforce account [name] — all opps, contacts, recent activity:
1. Build interactive Sites command center:
- Pipeline (stage, amount, close date)
- 7d signals (email, meetings, tickets)
- Prioritized next actions
2. Scheduled Task: weekdays 08:00 refresh Site
3. Material change → @Slack DM
Constraints: no auto external email; amounts = CRM raw.4.3 C: Lead audit & pipeline repair (Zapier-adapted)
Analyze @Salesforce 30d new leads + follow-ups; cross-check @Gmail sales threads.
Find:
1. Leads >48h no follow-up (by source)
2. Follow-chain break (step where response cliff)
3. Estimated pipeline loss $
Output:
- Excel (lead ID | source | last touch | break type | action)
- 1-page exec PPT highlighting seven-figure opportunity loss
- Weekly Scheduled audit flow proposal5. Marketing: два scenario
5.1 A: Research → brief → multi-market assets
Uploaded research: [attach / @Drive link]
End-to-end marketing workflow:
Phase 1 — Brief: audience, pain, positioning → Campaign Brief (Docs) + message pillars + channels
Phase 2 — Assets: from brief → 1 acquisition email, 3 LinkedIn posts, 1 LP outline → @Drive Campaign/[product]
Phase 3 — Geo: US/EU/APAC versions (language, culture, compliance); flag sensitive lines for human review
Pause after each phase; proceed on confirm.5.2 B: Slack/Teams → weekly agenda cron
Every Monday 07:00:
1. Aggregate @Slack #product-launch + @Microsoft Teams GTM (7d)
2. Extract decisions, open questions, meeting blockers
3. Update @Drive "weekly agenda" doc (keep history)
4. Post ≤5 bullets to @Slack #leadership
Constraints: public threads only; no confidential leak.6. Finance: два scenario
6.1 A: Month-end variance (OpenAI-validated)
Internal benchmark: close + forecast from days → hours.
Support [month] budget variance analysis:
1. Pull @Drive Finance/Actuals and Finance/Forecast
2. New @Sheets workbook: dept actual vs forecast; flag >5% or >$50K; preserve source formulas
3. Narrative draft (Docs): revenue / COGS / OpEx drivers
4. 5–8 page mgmt PPT (charts, template style)
5. List 3 finance judgment points for human sign-off
Constraints: no source mutation; cite source cells.6.2 B: Invoice vs payment recon (AP gate)
AP specialist. Compare:
- Payment register [@Drive link]
- Invoice list [@Drive link]
Flag table:
| issue | vendor | invoice # | amount | action |
>2% variance, missing tax ID, duplicate invoice #, vendor name mismatch
No auto payment; review table only.7. Ops: два scenario
7.1 A: Daily dashboard delta cron
Weekdays 06:30:
1. Access [dashboard URL / @SharePoint report]
2. vs yesterday snapshot: >10% moves or new red KPIs
3. 1-page morning brief (Docs): TOP 3, metric table, owners
4. @Gmail to ops-leads@company.com
If dashboard unreachable — report in Plan; do not hallucinate metrics.7.2 B: Feedback clustering → product priority
Monitor 14d feedback from:
- @Slack #customer-feedback
- @Gmail label NPS-Detractor
- @Drive Support Tickets Export
1. Cluster into 5–8 themes (representative quotes)
2. Score freq × impact × effort
3. Product backlog doc (Notion/Docs)
4. Friday Scheduled refresh
Constraints: anonymize PII.8. Product: launch readiness (Nvidia-adapted)
Launch readiness for [product/feature]:
1. @Jira: Epic/Story completion + open blockers
2. @Drive GTM Plans: milestone check
3. @Slack #product-launch: 7d unresolved threads
4. Readiness report (Docs): R/Y/G score, blockers (owner|due|risk), Go/No-Go + rationale
No auto Jira status change; high-risk = human decision.9. Engineering: Work × Codex split
Codex = code graph; Work = cross-team docs — same desktop binary, mode switch без context loss на Apple Silicon при достаточной unified memory.
9.1 A: PR review + release notes
Codex mode:
1. Review [repo] PR #123 (security/perf/coverage focus)
2. Inline review comments
3. If pass → Release Notes draft
Work mode:
4. Format for @Confluence
5. Draft @Slack #engineering announcement (no auto-send)9.2 B: Multi-repo weekly rollup
Codex across [frontend-repo] + [backend-repo]:
1. Merged PRs + open P0/P1 issues
2. Engineering weekly Markdown
Work:
3. Convert to Google Docs + burn-down from @Jira
4. Scheduled Task: Friday 17:0010. Scheduled Tasks recipe table
| Recipe | Trigger | Pipeline | Role |
|---|---|---|---|
| Monday agenda | Mon 07:00 | Slack → agenda doc | Marketing / Ops |
| Daily KPI brief | Weekdays 06:30 | Dashboard delta → email | Ops / Finance |
| Feedback cluster weekly | Fri 16:00 | Multi-channel → themes → priority | Product |
| Account daily refresh | Weekdays 08:00 | CRM delta → Sites hub | Sales |
10.1 Scheduled Task prompt pattern
Configure Scheduled Task:
- frequency: [daily / weekly Mon / 1st of month / @Slack keyword]
- time: [TZ + clock]
- action: [workflow description]
- notify: [Slack / email / none]
- human approve: [steps requiring pre-approval]10.2 Pre-unattended safety checklist
- Plugin scope minimized
- External auto-send OFF unless explicit
- Fixed output path — no overwrite
- Enterprise: Admin agent network policy verified
- 2–3 single-run validates before cron
11. Consumption optimization: до 5× delta
Work и Codex делят единый usage pool. Long context pulls раздувают effective memory footprint agent run — billing растёт нелинейно.
11.1 Billing drivers
| Factor | Consumption impact |
|---|---|
| Step count | More steps → higher burn |
| Context size | More docs/mail in window → ↑ tokens |
| Output length | Output ~6× input token cost |
| Cache hit | Repeat read ~1/10 fresh input |
| Model tier | GPT-5.6 reasoning premium vs light tasks |
11.2 Seven cost cuts
- Chat draft → trimmed prompt to Work
- Plan prune duplicate source pulls
- Scheduled template reuse for cache discount
- Concise output spec — table + 3 bullets
- Phase split direction → deliverable
- Free tier: benchmark small desktop jobs first
- Enterprise: workspace/group/user quotas in Admin Console
11.3 Pre-prod usage trial (5 steps)
- Pick task with known manual duration (e.g. variance table ~2h)
- One Work + Plan run; log step count
- Compare consumption vs included usage
- Extrapolate daily/weekly monthly burn
- If high → apply §11.2 and re-benchmark
12. Troubleshooting matrix
| Symptom | Root cause | Fix |
|---|---|---|
| Codex project missing in Work | Migration incomplete | Update Codex App → ChatGPT desktop; else reinstall from chatgpt.com/download |
| OAuth OK, empty pull | Scope / @typo | Catalog scopes; explicit @Salesforce not «CRM» |
| Plan OK, output drift | Stale context inference | Pause; attach fresh data |
| Cron no fire | Sleep / offline host | Long jobs → Web Agent; desktop must stay awake |
| Usage spike | Long output, dup pulls | §11; Admin caps |
| Work vs Cowork | Different workload class | Cloud SaaS = Work; local folder batch = Cowork (sister article) |
13. 30-day rollout roadmap
| Phase | Goal | Action |
|---|---|---|
| Week 1 | Single-task fluency | Desktop Work ×3 manual runs + Plan approval drill |
| Week 2 | Plugin integration | Mail + collab + files; one E2E artifact |
| Week 3 | Automation | Week-1 task → Scheduled; validate 3 fires |
| Week 4 | Team scale | Role prompt library; Enterprise quota sync |
14. FAQ
Q: С какой роли начать?
A: Там, где вы сами валидируете output. OpenAI: variance, brief, sales prep.
Q: Длина prompt?
A: Source + format + constraints, 150–400 слов. Не SOP кликов.
Q: Cron при выключенном ПК?
A: Desktop требует online host. Unattended → Web Workspace Agent.
Q: Work vs Workspace Agent?
A: Personal agent vs team-governed automation.
Q: PPT/Excel в external reporting?
A: 80% draft; numbers/names/statements — human QA.
Q: Free tier?
A: Desktop trial с cap; start invoice recon.
Sources: OpenAI, Sales Meeting Prep Cookbook, Changelog, SiliconANGLE. Updated 2026-07-11.
15. Execution node: Scheduled Tasks на Mac 24/7 с Metal-throughput
ROI ChatGPT Work — delegation рутинных cross-app pipelines, не «ещё одна AI-фича». Fast path: familiar task → 3 manual tuning runs → Plan trust → Scheduled takeover. 12 role prompts + 4 cron recipes — copy-ready.
Hard constraint desktop layer: execution host must stay online, powered, low-jitter network. Laptop sleep или Wi-Fi drop убивает 06:30 KPI brief и 08:00 account refresh — это deployment bottleneck, не prompt quality. Computer Use на Apple Silicon дополнительно упирается в Metal throughput и unified memory: cron на throttled mobile CPU деградирует latency multi-step jobs.
Production cron (sales brief, month-end recon, eng weekly) разумно класть на 7×24 Apple Silicon Mac node с ChatGPT desktop, workspace sync через SFTP/rsync.SFTPMAC remote Mac rental — macOS execution surface для AI agents: native Apple stack, stable plugin callbacks, no-sleep uptime, predictable Metal performance vs «личный MacBook как runner».