2026 CHATGPT
WORK_
6 ROLES_
PIPELINE.
Вводная: 9 июля 2026 OpenAI выкатил ChatGPT Work. Практический вопрос: что деплоить в production завтра? Официальный entry point — задача, которую вы уже умеете верифицировать. Этот гид разбирает Sales, Marketing, Finance, Operations, Product, Engineering на воспроизводимые pipeline с prompt-шаблонами, Plan Mode audit checklist, Scheduled Tasks recipes и usage pool optimization. Launch context + Cowork diff: ChatGPT Work + Codex merge — технический разбор.
1. Root cause: знаете Work, но не эксплуатируете
- Mode mismatch → до 5× usage delta — trivial Q&A в Work вместо Chat сжигает shared consumption pool.
- Prompt описывает steps, не deliverable — Work строит execution path сам; вы задаёте output contract.
- Scheduled Tasks без security preflight — unrestricted plugin scope + auto-send = compliance incident waiting to happen.
2. Три принципа (non-negotiable)
| Принцип | Механика | Implementation |
|---|---|---|
| Outcome, не procedure | Autonomous path planning | ❌ «Open Salesforce, export…» → ✅ «Из @Salesforce deals 30d: weekly PPT с risk flags» |
| Plugin stack first | Plugin catalog = data layer | Gmail/Slack/Drive authorized; explicit @AppName — снижает false pulls |
| Plan Mode = circuit breaker | Plan-before-execute | External email, finance reports, customer deliverables: line-by-line approval |
2.1 Mode routing table: Chat / Work / Codex
| Requirement | Mode | Rationale |
|---|---|---|
| Fast Q&A, brainstorm, single-turn copy | Chat | Min latency, min token burn |
| Cross-app multi-step, deliverables, hour-long runs | Work | Plugin stack + Plan Mode + Computer Use |
| Code review, PR ops, multi-repo dev | Codex | Dev-native toolchain |
| Recurring unattended background jobs | Work + Scheduled Tasks | Cron/trigger без manual kickoff |
2.2 Desktop vs Web deployment matrix
| Scenario | Target env |
|---|---|
| Local files, Computer Use, Free-tier probe | Desktop (Mac / Windows) |
| Team progress monitoring, mobile check | Web / Mobile (Plus+) |
| Sales brief pipeline + email notify | Web Workspace Agent + scheduler |
| Excel reconciliation, folder batch | Desktop Work mode |
3. Universal 5-step execution pipeline
3.1 Work prompt contract formula
3.2 Plan Mode audit checklist
- Data source + time window корректны (client/month)?
- Plan содержит send/delete/overwrite — high-risk ops?
- Output schema matches team template?
- Redundant steps removable для token savings?
- Human-in-the-loop gate required?
4. Шесть ролей: workflow + prompt templates
Compiled from OpenAI cases, Zapier/Nvidia/Virgin Atlantic early adopters, Workspace Agent Cookbook. Swap @plugin под ваш stack.
4.1 Sales
Scenario A: Daily meeting brief (Scheduled) — Baseline: 1–2 h/day manual → target <15 min review.
OpenAI internal: Discovery → custom PoC за 24h vs weeks.
Scenario B: Account command center (Sites, daily refresh)
Scenario C: Lead audit & pipeline repair (Zapier-adapted)
4.2 Marketing
Scenario A: Research → Brief → multi-market assets (E2E)
Scenario B: Slack/Teams → weekly agenda (Monday 07:00)
4.3 Finance
Scenario A: Month-end variance (OpenAI-validated) — Days → hours close cycle.
Scenario B: Invoice/payment reconciliation (AP automation)
4.4 Operations
Scenario A: Daily dashboard delta monitor (06:30 cron)
Scenario B: Feedback clustering → product priority
4.5 Product
Scenario A: Jira + GTM launch readiness (Nvidia-adapted)
4.6 Engineering — Work + Codex split
Codex = code path. Work = cross-team docs. Same desktop app context switch.
Scenario A: PR review + release notes
Scenario B: Multi-repo weekly engineering report
5. Scheduled Tasks recipe library
| Recipe | Trigger | Pipeline | Role |
|---|---|---|---|
| Monday agenda refresh | Mon 07:00 | Slack digest → agenda doc | Marketing / Ops |
| Daily KPI brief | Weekday 06:30 | Dashboard delta → email | Ops / Finance |
| Feedback weekly cluster | Fri 16:00 | Multi-channel → themes → priority list | Product |
| Account pulse | Weekday 08:00 | CRM delta → Sites update | Sales |
5.1 Scheduled Task prompt syntax
5.2 Unattended preflight security checklist
- Plugin scope minimized (least privilege)
- Auto-send disabled unless explicitly required
- Fixed archive path — no overwrite of shared assets
- Enterprise: agent network policy verified in Admin Console
- 2–3 manual dry-runs before cron activation
6. Usage optimization: consumption pool economics
Work и Codex share единый usage billing pool. Identical workflow, different design: до 5× delta.
| Factor | Consumption impact |
|---|---|
| Step count | Linear positive correlation |
| Context size | More docs/emails = more input tokens |
| Output length | Output tokens ~6× input cost |
| Cache hit | Cached input ~1/10 of fresh read |
| Model tier | GPT-5.6 heavy reasoning vs lightweight tasks |
6.1 Семь hardcore cost-cut tactics
- Chat draft → Work execute — prompt size −30–50%
- Plan Mode: strip redundant data pulls
- Scheduled Tasks: reuse template doc (cache discount)
- Compact output spec: «table + 3 bullets» > narrative report
- Phase split: Phase 1 direction lock → Phase 2 deliverable gen
- Free tier: probe small desktop tasks, log consumption before scale
- Enterprise: workspace/group/user caps in Admin Console
6.2 Pre-deploy usage estimation (5 steps)
7. Troubleshooting matrix
| Symptom | Root cause | Fix |
|---|---|---|
| Work can't find Codex project | Incomplete app migration | Update Codex → ChatGPT Desktop; fallback chatgpt.com/download |
| Plugin authorized, zero data | Scope / @ typo | Check plugin dir; @Salesforce not generic «CRM» |
| Plan OK, output drift | Stale context / inference | Pause + steer; attach critical data explicitly |
| Scheduled Task no-fire | Sleep / desktop logged out | Long runs: Web Agent; keep desktop awake |
| Usage > forecast | Long output, duplicate pulls | §6 optimization; Admin Console limits |
| Work vs Cowork? | Different workflow classes | SaaS cross-app = Work; local folder batch = Cowork |
8. 30-day onboarding roadmap
| Phase | Target | Action |
|---|---|---|
| Week 1 | Single-task proficiency | 1 task, desktop Work, 3 runs, Plan audit drill |
| Week 2 | Plugin deep integration | 3 core tools; 1 E2E deliverable |
| Week 3 | Automation | Week-1 task → Scheduled; verify 3 triggers |
| Week 4 | Team rollout | Role prompt library; Enterprise quota config |
9. Case study: Virgin Atlantic + OpenAI Finance validation
Virgin Atlantic сжал sales meeting prep с 1–2h manual до automated brief pipeline. OpenAI Finance ужал month-end variance с days до hours. Driver: repeatable process → standardized prompt + Scheduled Task — не «smarter model», а process engineering.
На Mac эти pipeline оптимальны в desktop Work mode (Computer Use для local Excel, FCP folders, Xcode simulator). При parallel Scheduled Tasks (sales brief + finance close + ops morning brief) unified memory и UI responsiveness становятся bottleneck — аргумент за tiered «Mac frontend + remote backend» architecture.
10. FAQ
Q: С какого role workflow стартовать?
Задача, которую вы быстрее всего валидируете на ошибки: variance analysis, marketing brief, sales meeting prep.
Q: Оптимальная длина prompt?
150–400 слов — data source, output format, constraints. Zero micro-step instructions.
Q: Scheduled Tasks при powered-off device?
Desktop cron требует online device. True 7×24: Web Workspace Agent (Plus+).
Q: Work vs Workspace Agent?
Work = personal agent mode. Workspace Agent = team automation с Admin governance.
Q: PPT/Excel direct external use?
80%-draft — numbers, names, claims require manual verification.
Q: Free tier — какие templates?
Lightweight tasks (invoice reconciliation) на desktop; no long-running cron until consumption profiled.
11. References
- OpenAI Blog — ChatGPT Work Launch
- OpenAI Cookbook — Sales Meeting Prep Agent
- ChatGPT Learn Changelog
- Связанный пост: ChatGPT Work launch + Cowork comparison
12. Conclusion: Mac local Work, long-running agents на MACGPU
ChatGPT Work окупается, когда убирает manual routine. Mac desktop — optimal host для Work + Codex + Computer Use: local Excel, folder batch, GUI automation на Apple Silicon. Но parallel Scheduled Tasks + multi-hour agent runs + multi-repo Codex съедают unified memory и battery; Windows cloud hosts не дают native Apple toolchain.
Production-grade split: Mac local — Work interaction + Plan Mode approval; sales brief pipeline, finance close, ops morning brief как Scheduled agents на MACGPU Remote Mac mini M4 — Apple Silicon unified memory для parallel runs, SSH on-demand, «frontend control + backend resident» tiered architecture.