ChatGPT Work на практике:
6 role-based workflow, prompt-шаблоны и automation (2026)

Если вы sales, marketing, finance, ops, product или engineering и после релиза 9 июля 2026 спрашиваете, что делать с ChatGPT Work в понедельник утром — этот технический гайд следует совету OpenAI: начать с задачи, которую вы уже знаете. Внутри — копируемые prompts, checklist Plan Mode, четыре рецепта Scheduled Tasks, семь тактик оптимизации consumption, 30-дневный roadmap и FAQ. Контекст релиза и сравнение с Claude Cowork: разбор запуска ChatGPT Work.

[ SECTION_01 ] // PRINCIPLES Перед копированием prompt: три принципа и выбор режима

Три принципа ChatGPT Work
Принцип Смысл Практика
Описывать результат Work сам планирует путь Не «открой Salesforce…», а «PPT pipeline из @Salesforce за 30 дней с risk flags»
Сначала подключить tools Plugins = data layer Gmail, Slack, Drive авторизованы; @AppName для источников
Plan Mode как тормоз План → approval → execution Внешние письма, finance reports, client deliverables — пошаговое OK
  • Неверный режим: Chat вместо Work для multi-app pipeline раздувает consumption.
  • Plugins не подключены: длинный prompt без CRM/mail = пустые данные или галлюцинации.
  • Plan Mode пропущен: external send, delete, overwrite без контроля.
  • Desktop vs Web: локальный Excel, Computer Use → desktop; team Scheduled Tasks → Web Workspace Agent.
Chat / Work / Codex — матрица выбора
Задача Режим Почему
Быстрый Q&A, один текст Chat Лёгкий runtime
Multi-app, deliverables, часы Work Plugins + Plan Mode + Computer Use
Code review, PR, multi-repo Codex Dev-native workflow
Фоновая automation Work + Scheduled Tasks Schedule/trigger execution

OpenAI рекомендует: month-end variance, campaign brief или prep sales meeting — задачи, качество которых вы можете проверить сразу.

[ SECTION_02 ] // FRAMEWORK Универсальный 5-шаговый workflow и формула prompt

WORKFLOW.txt 
1. Подключить plugins
2. Цель + output format
3. Review Plan Mode
4. Mid-flight correction
5. Accept deliverable & iterate
PROMPT-FORMULA.txt 
[Role] + [Data sources @plugins] + [Task] + [Output format] + [Constraints] + [Acceptance criteria]

Plan Mode checklist перед execution:

  • Источники, период и account верны?
  • External send, delete, overwrite?
  • Output совпадает с team template?
  • Redundant data pulls можно убрать?
  • Нужен human approval checkpoint?

[ SECTION_03 ] // ROLES Шесть role-based workflow с prompt-шаблонами

Шаблоны на базе кейсов OpenAI, feedback Zapier/Nvidia/Virgin Atlantic и Workspace Agent Cookbook. Замените @plugin на ваш stack.

Sales — A: Daily meeting briefs (Scheduled)

SALES-A-SCHEDULED.txt 
Scheduled Task будни 16:00:
1. Customer meetings завтра @Google Calendar (internal exclude)
2. 30-day notes @Salesforce/@SharePoint, company news 30d
3. 2–3 предложения background на external attendee
4. Brief 2–3 стр/meeting → @Google Drive
5. Summary @Gmail (таблица: client|time|topic|link)

Sales — B: Account command center (Sites)

SALES-B-SITES.txt 
@Salesforce [Account]: interactive Sites dashboard
Pipeline, 7-day signals, prioritized next actions
Refresh 08:00 weekdays; @Slack DM on major changes
No external emails; amounts from CRM raw data

Sales — C: Lead review & pipeline repair

SALES-C-PIPELINE.txt 
@Salesforce leads 30d + @Gmail outreach:
48h+ follow-up gaps, broken handoffs, estimated pipeline loss
Excel detail + 1-page executive PPT + weekly Scheduled review flow

Marketing — A: Research → Brief → multi-market assets

MKT-A-ENDTOEND.txt 
Phase 1: Campaign Brief Docs (audience, pain, messaging pillars)
Phase 2: acquisition email, 3 LinkedIn posts, landing outline → Drive Campaign/[product]
Phase 3: US/EU/APAC localization notes; pause after each phase

Marketing — B: Slack/Teams → weekly agenda

MKT-B-AGENDA.txt 
Mondays 07:00: @Slack #product-launch + @Teams GTM, 7 days
Decisions, open items, blockers → Drive weekly agenda doc
≤5 bullets @Slack #leadership; no confidential content

Finance — A: Month-end variance (OpenAI-validated)

FIN-A-VARIANCE.txt 
[Month] budget variance:
@Drive Finance/Actuals + Forecast → @Sheets (>5% or >$50K flag)
Narrative Docs, 5–8 slide deck, 3 human judgment calls
Source files untouched; cite cell references

Finance — B: Invoice vs payment reconciliation

FIN-B-AP.txt 
Payment register vs invoice list @Drive:
>2% amount diff, missing tax ID, duplicate invoice, vendor mismatch → review table
Do not initiate payments

Ops — A: Daily dashboard briefing (Scheduled)

OPS-A-DASHBOARD.txt 
Weekdays 06:30: [dashboard URL/@SharePoint] vs yesterday snapshot
>10% swings → 1-page Docs → @Gmail ops-leads
If unreachable: stop in Plan phase — do not fabricate data

Ops — B: Feedback clustering → product priorities

OPS-B-FEEDBACK.txt 
14 days: @Slack #customer-feedback, @Gmail NPS-Detractor, Drive tickets
5–8 themes, rank frequency×impact×effort, Friday Scheduled refresh, anonymized

Product — Launch readiness (Jira + GTM)

PROD-A-READINESS.txt 
[Feature]: @Jira completion/blockers + @Drive GTM milestones + @Slack #product-launch
Red/Yellow/Green readiness, Go/No-Go; do not auto-update Jira

Engineering — Codex + Work collaboration

ENG-A-PR-RELEASE.txt 
Codex: review PR #123 [repo], sidebar comments, release notes draft
Work: @Confluence format, @Slack #engineering post draft (no auto-send)
ENG-B-WEEKLY.txt 
Codex: [frontend]+[backend] merged PRs, open P0/P1 → Markdown weekly
Work: Google Docs + @Jira burndown, Fridays 17:00 Scheduled

[ SECTION_04 ] // AUTOMATION Scheduled Tasks, security checklist и usage optimization

Четыре рецепта Scheduled Tasks (OpenAI)
Рецепт Trigger Action Роль
Monday agenda Mon 07:00 Slack digest → agenda doc Marketing/Ops
Daily metrics brief Weekdays 06:30 Dashboard diff → email Ops/Finance
Feedback clustering Fri 16:00 Multi-channel → priority list Product
Account daily refresh Weekdays 08:00 CRM → Sites dashboard Sales
SCHEDULED-TASK.txt 
Scheduled Task:
- Frequency: [daily/Monday/1st of month/@Slack keyword]
- Time: [timezone + time]
- Action: [workflow description]
- Notification: [Slack/email/none]
- Human approval: [steps requiring pre-OK]

Security checklist перед unattended mode: minimal plugin scope, no auto-external-send, fixed archive path, Enterprise agent network policy confirmed, 2–3 manual runs before schedule.

Факторы consumption billing
Фактор Влияние
Step count Больше шагов = выше consumption
Context size Volume docs/emails pulled
Output length Output tokens ≈ 6× input cost
Cache hit Repeat read ≈ 1/10 fresh input
Model GPT-5.6 vs lightweight tasks
  • Draft в Chat, compact brief в Work
  • Plan Mode: trim duplicate pulls
  • Template docs в Scheduled Tasks для cache discount
  • Concise output (table + 3 bullets)
  • Phase split против expensive re-runs
  • Free tier: small desktop tests before scale
  • Enterprise: workspace/group/individual limits в Admin Console

[ SECTION_05 ] // ROADMAP 30-дневный onboarding roadmap и 6 шагов старта

30-day ChatGPT Work onboarding
Неделя Цель Действие
1 Single-task fluency 3 manual Work runs, Plan Mode practice
2 Plugin depth 3 core tools, 1 cross-app deliverable
3 Automation Week 1 task → Scheduled, verify 3 triggers
4 Team rollout Role prompt library, admin limits
  1. Установить desktop: chatgpt.com/download; обновить Codex app если есть.
  2. Work mode: авторизовать plugins Gmail, Slack, Drive.
  3. Первый prompt: скопировать шаблон §3, указать @app, format, constraints.
  4. Plan Mode: checklist §2, затем execution.
  5. Accept deliverable: 80% draft; numbers, names, external statements — manual review.
  6. Scheduled Task: после 2–3 успешных runs; Web Agent с Plus+.
USAGE-TEST.txt 
1. Real task with known human time cost
2. One Work + Plan Mode run, count steps
3. Consumption vs included usage
4. Extrapolate daily/weekly/monthly
5. Optimize per §4 and re-run compare

[ SECTION_06 ] // FAQ FAQ, pitfalls, citeable data и NOVAKVM

Typical pitfalls
Проблема Причина Fix
Codex projects missing Incomplete app migration Update Codex → ChatGPT Desktop; reinstall if broken
Plugin connected, no data Insufficient scope / wrong @name Re-check permissions; explicit @Salesforce
Good plan, wrong output Stale context / AI inference Pause and steer; attach source files
Scheduled task missed Device asleep / logged out Web Workspace Agents; desktop must stay online
Usage higher than expected Verbose output, redundant pulls §4 optimization; Admin Console limits
Work vs Cowork Different workflow types Cloud SaaS → Work; local folders → Cowork (sister post)
  • Какой workflow первым? Задача, которую вы знаете и можете verify — variance, brief, meeting prep.
  • Длина prompt? 150–400 слов: sources, format, constraints — не micromanage steps.
  • Scheduled при выключенном laptop? Desktop tasks need device online; true background → Web Workspace Agents (Plus+).
  • Work vs Workspace Agent? Work = personal agent mode; Workspace Agent = team-built automations in Business/Enterprise.
  • Slides/reports наружу as-is? 80% draft; human-review numbers, names, external statements.
  • Free users? Desktop Work with limits; start lightweight like invoice reconciliation.
  • Output token cost: примерно 6× input — таблицы выгоднее narrative.
  • Cache discount: repeat read ≈ 1/10 fresh input.
  • Cost spread: один workflow до 5× разницы по design.

Primary sources — reopen after product updates:

OpenAI Blog: ChatGPT for Your Most Ambitious Work

OpenAI Cookbook: Sales Meeting Prep Agent

ChatGPT Learn Changelog

SiliconANGLE: ChatGPT Work Launch

Developers Digest: Codex Merge Analysis

ChatGPT Work снимает ручную рутину, но Scheduled Tasks на sleeping laptop, expired OAuth и full disk сводят на нет выгоду 1400+ plugins. Pure cloud SaaS не заменяет stable local execution surface для Codex multi-repo development и Xcode validation на той же Apple Silicon machine.

Чтобы закрепить ChatGPT Work, длинные Codex runs и iOS CI/CD на фиксированном Apple Silicon host 7×24, перенесите Computer Use и GitHub PR review на dedicated bare metal вместо unstable endpoint. NOVAKVM — Mac Mini M4/M4 Pro, fixed bandwidth, SSH by default, оптимально для AI agent automation. Тарифы, Заказ, Центр помощи.