The Codex Computer Use Playbook — 10 Automation Prompts for Windows Desktop Tasks

Codex Computer Use on Windows: A 10‑Prompt Automation Playbook

Author: Markos Symeonides | Published: July 17, 2026

Codex Computer Use turns natural language instructions into concrete desktop actions, reliably automating what a human would do on a Windows machine: opening apps, navigating UI elements, reading on‑screen content, inputting data, and verifying outcomes. This playbook delivers 10 production‑grade prompts for Windows that you can paste directly into Codex to automate high‑value workflows across Outlook, Excel, browsers, file systems, accounting tools, social platforms, and CRM pipelines. Each prompt is backed by prerequisites, verification steps, troubleshooting, and extensions you can adapt for your environment.

All scenarios were designed and tested on Windows 11 (23H2/24H2), Microsoft 365 Apps for enterprise (Current Channel, Version 2409+), Microsoft Edge (Stable 126+), and common line‑of‑business apps. Where a product version matters, we call it out explicitly. If you use a different Windows build or application version, adjust selectors and navigation accordingly. For broader context on architectural choices and strategy, see our companion explainers:

Understanding the cost implications of these features requires familiarity with OpenAI’s current tier structure. Our detailed breakdown in 5 Best AI Research Tools for writing Compared u2014 Features, Pricing, Use Cases compares every subscription option, usage limits, and value proposition to help you choose the right plan for your workflow.

,

Enterprise teams deploying these capabilities at scale must consider governance and compliance requirements. Our guide on 35 ChatGPT-5.6 Work Prompts for Enterprise Automation Connectors covers the essential security frameworks and access controls needed for production AI deployments.

,

For additional context on related AI capabilities and workflows, our comprehensive resource on The Big AI Coding Agents Story: What July 16’s News Means for Developers provides practical guidance and implementation strategies that complement the techniques discussed in this article.

, and

Understanding how these capabilities compare across different AI platforms provides valuable context for tool selection. Our head-to-head comparison in Claude Opus 4.7 vs OpenAI Codex for Indie Shipping: Which Should You Choose in 2026? evaluates the strengths and trade-offs of competing approaches.

.

Executive summary: What Computer Use enables and why it matters

Computer Use extends code generation with grounded, visible actions on a real desktop. Unlike pure API automation, it can operate any Windows application that a human can operate, even when no formal API or CLI exists. It observes the screen, reasons about the interface, performs actions (clicks, keystrokes, file operations, application launches), and checks the results before moving on. The result is a unified automation layer that spans legacy UIs, modern SaaS in the browser, Microsoft 365 apps, and internal tools.

• Automate cross‑app workflows that previously required brittle, app‑specific scripts.
• Reduce manual processing load in shared services: finance ops, IT ops, sales ops, research, QA.
• Close the “last mile” between data analysis and business action (email, tickets, postings, approvals).
• Standardize repeatable tasks with verification gates to keep humans in control of outcomes.

For leaders, Computer Use drives measurable gains in cycle time and error rate without waiting for APIs or vendor integrations. For practitioners, it provides a tactical instrument to ship reliable automations quickly and maintain them with prompt‑level diffs instead of large code refactors.

How Computer Use works technically (vision, action, verification loop)

At its core, Computer Use coordinates three capabilities in a tight loop:

1. Visual perception: The agent captures screenshots, interprets text via OCR, recognizes UI affordances (buttons, menus, tables, input fields), and parses layout structure. It uses past observations to remain context‑aware across steps (e.g., a modal obscuring the main window).
2. Action execution: The agent issues UI‑level commands—launching applications, typing, pasting, selecting menu items, scrolling, resizing windows, and interacting with system dialogs. On Windows, this includes Start Menu search, Taskbar pinning, File Explorer navigation, notification toasts, and common Win32/WinUI patterns.
3. Verification: After each action (or group of actions), the agent checks for expected textual or visual cues. If verification fails, it adapts: retrying with backoff, choosing alternative UI paths, or asking for clarification. Verification can be strengthened with explicit instructions and “must‑see” conditions defined in your prompt.

This loop is orchestrated by a planner that aligns each subgoal to a verifiable observation. Good prompts give the planner structure: naming target windows, expected labels, error messages to watch for, file paths, and acceptance criteria. The stronger your “definition of done,” the fewer surprises in production.

Why prompts matter as much as code

Unlike traditional RPA where selectors are hard‑coded, Computer Use relies on task descriptions plus dynamic perception. Your prompt defines the task contract: allowed applications, data sources, frequencies, and what counts as success. This enables robust generalization across minor UI changes while maintaining guardrails for sensitive actions.

Computer Use vs. traditional approaches

Setting up your environment for these prompts

Baseline system

• Windows 11 23H2 or 24H2 with display scaling at 100% or 125% (avoid unusual DPI that hides labels).
• Microsoft 365 Apps for enterprise (Outlook/Excel/Word). Tested on Current Channel Version 2409+.
• Microsoft Edge Stable 126+ or Google Chrome 126+ with English UI (adjust prompts if localized).
• File Explorer set to show file extensions and hidden items when dealing with file types.
• Stable network (wired or strong Wi‑Fi), and disable sleep during long automations.

Profiles, folders, and sample data

• Create C:\Ops\InboxRules\, C:\Ops\Reports\, C:\Ops\Backups\, C:\Ops\Temp\.
• Standardize downloads to C:\Ops\Downloads\ via your browser’s settings.
• For prompts involving credentials, store them in your password manager and rely on Windows Hello prompts; avoid hardcoding secrets in prompts.

Windows and app configurations to pin down

• Pin Outlook, Excel, Edge, File Explorer to the Taskbar for predictable launching.
• Ensure Outlook is connected to your Microsoft 365 tenant and that cached mode is on for speed.
• In Excel, enable trusted locations for C:\Ops\Reports\ to avoid macro/security popups when opening CSV/XLSX.

Artifacts you might need

# Example: Create standard directories (run in PowerShell as admin if needed)
New-Item -ItemType Directory -Force -Path "C:\Ops\InboxRules","C:\Ops\Reports","C:\Ops\Backups","C:\Ops\Temp","C:\Ops\Downloads" | Out-Null

# Example: Windows sleep policy for logged-in session (set display off but no sleep)
powercfg -change -standby-timeout-ac 0
powercfg -change -monitor-timeout-ac 20

Safety and permissions considerations

• Principle of least privilege: run on a user profile with only the drives and apps needed for the task.
• Consent gating: require explicit confirmation in prompts before sending external communications or deleting files.
• Data segregation: stage working files under C:\Ops\Temp\ and move to destination only after verification passes.
• Logs: instruct the agent to export an activity log (timestamps and screenshots) to C:\Ops\Backups\Logs\.
• PII handling: redact personally identifiable info in screenshots stored for logs or use a secure log location.
• Dry‑run mode: for high‑impact tasks, include a switch to simulate clicks and collect screenshots before enabling writes.

Ten production‑grade prompts for Codex Computer Use on Windows

Each prompt below is ready to paste into Codex. Tailor file paths, email domains, and app names for your environment. To harden reliability, keep labels literal (e.g., “Reply All”) and include fallback cues (e.g., “if not visible, open View > Reading Pane”).

1) Automated Email Triage and Response Drafting (Outlook)

Complete prompt

You are automating Outlook triage on Windows 11 using Computer Use. Follow these rules:
Goal:
- Triage today's unread messages in the Inbox from 06:00 to now.
- Categorize into: "Invoices", "Meetings", "Support", "Other".
- Draft responses using approved templates; do not send automatically without my confirmation.
- Create a summary table (Sender, Subject, Category, Proposed Action) and save to C:\Ops\Reports\Inbox_Triage_[YYYY-MM-DD].xlsx.

Applications you may use:
- Outlook for Microsoft 365 (desktop)
- Excel (desktop)
- Notepad
- File Explorer

Step-by-step:
1) Launch Outlook. Go to Inbox. Filter to Unread and "Received Today".
2) For each unread message (limit 50):
   - Open message in Reading Pane; if off, turn on View > Reading Pane > Right.
   - Extract Sender, Subject, Received time, first 300 chars of body.
   - Categorize:
       * "Invoices" if subject/body contains "invoice", "bill", "PO", "payment due".
       * "Meetings" if contains "meeting", "invite", "calendar", "reschedule".
       * "Support" if contains "issue", "ticket", "error", "bug", "urgent help".
       * Else "Other".
   - For "Invoices": reply with a draft:
       Subject prefix: "Re: " + original subject
       Body template:
         Hi [Sender First Name],
         Thank you for the invoice. Please confirm the invoice number and due date. We will process within 2 business days.
         Regards,
         [Your Name]
     For "Meetings": propose 3 time slots in the next 3 business days, 30-min duration, between 09:00-16:00 local time.
     For "Support": acknowledge receipt and ask for logs, screenshots, and steps to reproduce.
     For "Other": acknowledge receipt and state response within 2 business days.
   - Save each draft (do not send). Ensure "To" is pre-populated; if not, populate from the message's sender.
   - Add Outlook Category with the category name.
3) Build an Excel workbook with a single sheet "Triage". Columns: Sender, Subject, Category, Proposed Action, DraftPath (if saved as .msg), LinkToMessage (Outlook item link if available).
4) Save the Excel file to C:\Ops\Reports\Inbox_Triage_[YYYY-MM-DD].xlsx. Overwrite if exists.
5) Present a final confirmation dialog with counts per category and "Send all drafts now?" with Yes/No buttons. If I click Yes, send drafts for "Invoices" and "Support" only; leave others unsent.

Verification:
- After categorizing each message, confirm that the Outlook Category pill is applied.
- For each draft, verify the Subject begins with "Re:" and the body contains the correct template section.
- Verify Excel file exists and has row count equal to number of messages processed.
- Before sending, display the confirmation and wait for my input.

Constraints:
- Do not delete or archive any emails.
- Stop after 50 messages or when no more unread messages from today.
- If Outlook search or filter UI differs, use Search box: received:today isread:no.
- Log actions with timestamps to C:\Ops\Backups\Logs\InboxTriage_[YYYY-MM-DD].txt.

If an error occurs:
- Take a screenshot, write an entry to the log, skip the message, and continue.
- If Outlook is not signed in, stop and ask me to sign in.

Prerequisites and setup steps

• Outlook for Microsoft 365 configured and cached. Ensure Reading Pane is available (View tab).
• Excel installed and not blocked by protected view for C:\Ops\Reports\.
• Create C:\Ops\Backups\Logs\ in advance.
• Outlook Categories pre‑created: Invoices, Meetings, Support, Other.

Expected behavior and verification

• Unread messages from today are categorized and drafts are prepared without sending.
• An Excel file summarizes the triage with correct row counts.
• Confirmation dialog prevents accidental sending; only specified categories are sent if approved.

Troubleshooting tips

• If Outlook’s Filter Email button is hidden, use the Search bar query “received:today isread:no”.
• If protected view blocks Excel writes, add C:\Ops\Reports\ as a Trusted Location (Excel Options > Trust Center).
• If categories aren’t visible, enable the Categories column in the view or apply via Message > Tags > Categorize.

Variations and extensions

• Route “Invoices” to a subfolder after draft creation.
• Attach a standardized PDF policy to “Support” replies.
• Append a footer with your SLA linked to

  Understanding the cost implications of these features requires familiarity with OpenAI’s current tier structure. Our detailed breakdown in 5 Best AI Research Tools for writing Compared u2014 Features, Pricing, Use Cases compares every subscription option, usage limits, and value proposition to help you choose the right plan for your workflow.

  .

2) Excel Report Generation from Multiple Data Sources

Complete prompt

You are automating a daily Excel report assembly on Windows 11.

Goal:
- Aggregate data from CSV, a SharePoint-hosted XLSX, and a downloaded HTML table.
- Clean and combine into a single workbook with pivot and chart.
- Save to C:\Ops\Reports\Daily_Ops_[YYYY-MM-DD].xlsx and export a PDF.

Applications allowed:
- Excel (desktop)
- Microsoft Edge
- File Explorer
- PowerShell (optional via Windows Terminal)

Data sources:
- Local CSV: C:\Ops\Downloads\sales_[YYYYMMDD].csv
- SharePoint XLSX: https://contoso.sharepoint.com/sites/BI/Shared%20Documents/Inventory.xlsx
- Web table: https://example.com/market.html (table with id "prices")

Steps:
1) Launch Edge. Download the web table as CSV by copying the HTML table into Excel, or use "Copy" then paste into Excel and Save As CSV in C:\Ops\Temp\market_[YYYYMMDD].csv.
2) Open Inventory.xlsx in the browser; choose "Open in Desktop App".
3) Open Excel and import:
   - From Text/CSV (sales_*.csv)
   - From Workbook (Inventory.xlsx)
   - From Text/CSV (market_*.csv)
4) Use Power Query to:
   - Ensure date column types are Date, numeric columns as Decimal.
   - Remove rows with null key fields.
   - Standardize column names: Region, SKU, Qty, Price, Date.
   - Merge sales with inventory on SKU (left join) and market on SKU (left join).
5) Load to a new worksheet "DataModel". Create a PivotTable in "Summary":
   - Rows: Region
   - Values: Sum of Qty, Sum of Revenue (=Qty*Price)
6) Insert a clustered column chart for Revenue by Region, title "Revenue by Region – [YYYY-MM-DD]".
7) Save workbook to C:\Ops\Reports\Daily_Ops_[YYYY-MM-DD].xlsx (overwrite ok).
8) Export PDF of "Summary" to C:\Ops\Reports\Daily_Ops_[YYYY-MM-DD].pdf.
9) Close all without saving temporary external files.

Verification:
- The Summary sheet exists, pivot totals are nonzero, and chart title contains today's date.
- Row count in DataModel > 50 (sanity check).
- PDF exists and is readable.

Error handling:
- If SharePoint "Open in Desktop App" fails, download and open locally.
- If Power Query errors, open "Queries & Connections" and fix data types; retry refresh.
- Log steps to C:\Ops\Backups\Logs\DailyOps_[YYYY-MM-DD].txt.

Prerequisites and setup steps

• Excel with Power Query and permission to access the SharePoint site.
• Known browser downloads folder mapped to C:\Ops\Downloads\.
• Inventory.xlsx columns include SKU and QtyOnHand; sales CSV includes Date, SKU, Qty, Price.

Expected behavior and verification

• Consolidated DataModel with merged datasets.
• Pivot and chart created with correct measures and date title.
• PDF export of the Summary worksheet present in Reports.

Troubleshooting tips

• If SharePoint opens in the web app only, use File > Save As > Download a Copy, then open locally.
• If column headers vary, add a mapping step in Power Query to rename mismatches.
• If Excel Protected View blocks external files, trust the folder or unblock individual files.

Variations and extensions

• Append new daily data to a year‑to‑date workbook with a Refresh All macro button.
• Publish the PDF to Teams via the desktop or web client UI.

3) Browser‑Based Research and Data Collection

Complete prompt

Automate research across web sources and output a structured CSV.

Goal:
- Search 5 specified queries.
- For each top 10 result (excluding ads), collect Title, URL, Snippet, Publication Date if visible, and Organization.
- Deduplicate by URL host and title similarity; keep the most recent.
- Save to C:\Ops\Reports\Research_[YYYY-MM-DD].csv.

Applications:
- Microsoft Edge
- Notepad or Excel
- File Explorer

Inputs:
- Queries:
  1) "Windows 11 policy-based management best practices"
  2) "Outlook desktop automation troubleshooting"
  3) "Excel Power Query performance tips"
  4) "Selenium vs Computer Use for UI tests"
  5) "Invoice processing automation accounting"

Steps:
1) Open Edge. For each query, search on Bing and Google (two tabs each).
2) Scroll the SERP; ignore "Ad" or "Sponsored" results. Open top 10 organic links in background tabs.
3) On each result page:
   - Extract title (H1 or document title), URL, visible snippet/summary (first paragraph), and date if near the title or byline.
   - If a paywall blocks content, skip with note "paywalled".
   - Normalize organization using domain (e.g., "contoso.com" → "Contoso").
4) Build a list in memory; deduplicate:
   - Drop entries with same URL.
   - If titles are 85% similar via simple ratio (approximate by text comparison), keep the one with a visible date; else keep the first.
5) Save to C:\Ops\Reports\Research_[YYYY-MM-DD].csv with header: Query, Title, URL, Snippet, Date, Organization.
6) Summarize counts by query in a Notepad report and save to C:\Ops\Backups\Logs\Research_Summary_[YYYY-MM-DD].txt.

Verification:
- CSV exists with >= 20 rows and exactly 6 columns.
- At least 1 result per query.
- No lines containing "Ad" or "Sponsored" in Title.

Constraints:
- Do not sign in to any site.
- Respect robots and site policies; no rapid-fire refresh loops.

If blocked:
- Slow down scrolling, wait 5–10 seconds between pages, and continue.

Prerequisites and setup steps

• Edge with default search engines enabled; pop‑up blocker standard settings.
• Time zone correct to capture publication dates accurately.

Expected behavior and verification

• CSV populated with diverse, non‑ad results across both search engines.
• Deduplication reduces near‑duplicates and keeps dated articles when available.

Troubleshooting tips

• If cookie/consent banners obscure content, accept or close them before scraping text.
• If copy/paste introduces line breaks, paste into Notepad first, then into Excel/CSV for clean commas.

Variations and extensions

• Add a second pass to capture author names where visible.
• Create an HTML report with links and dates using Excel’s “Publish” or Notepad templating.

4) File Organization and Backup Automation

Complete prompt

Automate weekly file organization and backup to a dated archive.

Goal:
- From C:\Users\[User]\Downloads and C:\Ops\Temp, move files into C:\Ops\Backups\[YYYY-MM-DD]\ by type:
  - Documents: .pdf, .docx, .pptx, .xlsx, .csv → Docs\
  - Images: .png, .jpg, .svg → Images\
  - Archives: .zip, .7z, .rar → Archives\
  - Installers: .msi, .exe → Installers\
  - Other → Misc\
- Generate a manifest CSV with file name, source path, destination path, size (bytes), and SHA256.
- Compress the archive folder to C:\Ops\Backups\[YYYY-MM-DD].zip.

Applications:
- File Explorer
- Windows Terminal (PowerShell)
- Notepad

Steps:
1) Create C:\Ops\Backups\[YYYY-MM-DD]\ subfolders as listed.
2) Move files from Downloads and C:\Ops\Temp\; skip if in use (log and continue).
3) Generate manifest:
   - For each moved file, compute SHA256 and append a line to C:\Ops\Backups\[YYYY-MM-DD]\manifest.csv.
4) Compress the dated folder into a zip archive.
5) Leave the original dated folder in place for inspection.

Verification:
- Manifest exists and row count equals number of moved files.
- ZIP exists and is smaller than sum of file sizes when compression is expected (.text, .csv).
- Spot-check a few SHA256 hashes by recomputing.

Constraints:
- Never delete files; only move and copy/compress.
- If a filename collision occurs, append " (n)" before extension.

Logging:
- Write to C:\Ops\Backups\Logs\Backup_[YYYY-MM-DD].txt with start/end timestamps.

PowerShell helper (optional, you may run it):
- Use Get-FileHash -Algorithm SHA256 and Compress-Archive.

Prerequisites and setup steps

• Ensure you have space on C: drive for compression.
• PowerShell 5.1+ available (Windows 11 includes this; PowerShell 7 optional).

Expected behavior and verification

• All files moved into a dated, typed folder structure with a verified manifest.
• A compressed ZIP archive is created for offsite copy.

Troubleshooting tips

• If Access Denied appears, re‑run File Explorer as the current user (not admin) to avoid UAC context differences.
• Long paths may fail; enable long paths in Group Policy or shorten directory depth.

Variations and extensions

• Mirror the ZIP to OneDrive using the OneDrive client UI to validate sync.
• Encrypt the ZIP with 7‑Zip via context menu instead of native compression.

5) Meeting Notes Extraction and Action Item Creation

Complete prompt

Automate extraction of meeting notes and creation of action items.

Goal:
- From the last 10 Outlook calendar events that have Teams meeting recordings and chat transcripts in the last 7 days:
  - Download transcript (.vtt or .docx) and meeting notes if present.
  - Extract key decisions and action items (Owner, Task, Due Date).
  - Create a consolidated Excel "ActionItems_[YYYY-MM-DD].xlsx" with a sheet per meeting and a master sheet.
  - Draft follow-up emails to owners listing their assigned items.

Applications:
- Outlook (desktop) and Teams (desktop or web)
- Excel (desktop)
- Edge (if Teams web used)
- Notepad

Steps:
1) In Outlook Calendar, filter past 7 days. For each meeting (limit 10):
   - Open the meeting, follow the "Chat" or "Meeting notes" link to Teams.
   - If a transcript is available, download it (VTT or DOCX) to C:\Ops\Temp\Meetings\[MeetingID]\.
   - If "Meeting notes" exist, copy content to Notepad and save in same folder.
2) Parse content:
   - Identify bullet points or lines starting with action verbs (e.g., "Send", "Draft", "Update").
   - Extract assignees by @mentions or nearest person name.
   - If a date phrase appears (“by Friday”, “EOD Thu”), normalize to a date within the next 14 days.
3) Build Excel:
   - One sheet per meeting with columns: Owner, Task, Due Date, Source (Transcript/Notes).
   - Master sheet "AllItems" aggregating all rows with Meeting Title, Date, Link.
4) Draft follow-ups:
   - For each person, draft one email listing all their tasks, grouped by meeting, due dates bolded.
   - Save drafts; do not send.
5) Save workbook to C:\Ops\Reports\ActionItems_[YYYY-MM-DD].xlsx.

Verification:
- Master sheet row count equals sum of per-meeting rows.
- Draft emails exist and list items grouped per owner.
- Each meeting folder contains at least a transcript or notes file.

Constraints:
- Respect privacy: do not upload files elsewhere.
- If no transcript exists, use chat summary.

Logging:
- Save a processing log per meeting under its folder and a global log under C:\Ops\Backups\Logs\Meetings_[YYYY-MM-DD].txt.

Prerequisites and setup steps

• Teams meeting content retention must allow transcript access for the last 7 days.
• Permissions to access meeting chats and notes via Outlook/Teams integration.

Expected behavior and verification

• Workbook with per‑meeting sheets and a consolidated master list of action items.
• Draft emails prepared per owner; no emails sent without approval.

Troubleshooting tips

• If Teams opens in the web and blocks downloads, use “Open in desktop app” or copy text from the viewer.
• If date normalization is ambiguous, leave Due Date blank and add a note “manual confirmation required”.

Variations and extensions

• Push “AllItems” into a Planner board via the web UI if allowed.
• Add a confidence column for each extracted item to prioritize review.

6) Software Testing: UI Regression Checks

Complete prompt

Automate a lightweight UI regression pass for a desktop app and its web portal.

Goal:
- Validate that key UI elements and user flows render correctly after a build.
- Capture screenshots on pass/fail and write a results HTML.

Target:
- Desktop app: "ContosoCRM" (WinUI)
- Web portal: https://portal.contoso-crm.internal/ (Chrome or Edge)

Applications:
- The desktop app (installed)
- Microsoft Edge
- Notepad

Test Plan:
1) Desktop app smoke tests (launch from Start menu):
   - Verify splash screen contains "ContosoCRM".
   - Login window: fields "Email" and "Password", and button "Sign in" visible.
   - Enter test credentials (use provided test account if cached; otherwise stop and request).
   - After login, verify tabs "Dashboard", "Accounts", "Opportunities".
   - Navigate to Accounts, open first row, verify "Edit" and "Save" buttons.
   - Take screenshots at: post-login dashboard, Accounts grid, Account detail.
2) Web portal smoke tests:
   - Open Edge, navigate to the portal URL.
   - Verify security banner and that the page title contains "Contoso CRM".
   - Login if required (use test account).
   - Navigate to /reports and verify a table with headers "Name", "Owner", "Updated".
   - Export a CSV via the "Export" button; verify file exists.
3) Compose an HTML results file with sections:
   - Environment details (Windows build, app version if visible)
   - Tests run with Pass/Fail
   - Inline thumbnails of screenshots linking to originals

Artifacts:
- Save screenshots to C:\Ops\Backups\UI\Run_[YYYYMMDD_HHMM]\.
- Results HTML at C:\Ops\Backups\UI\Run_[YYYYMMDD_HHMM]\index.html.

Verification:
- index.html loads locally and shows Pass/Fail counts.
- At least 5 screenshots exist for the run.
- The Export CSV exists in C:\Ops\Downloads\.

Constraints:
- Do not change production data.
- Use only test environment or read-only accounts.

Error handling:
- On failure, include the error text from the UI if available and the screenshot next to it.
- If login fails twice, abort remaining tests and mark as Blocked.

Prerequisites and setup steps

• Test accounts with read‑only roles.
• Access to the internal portal from the test machine (VPN if applicable).

Expected behavior and verification

• HTML report with screenshots and explicit pass/fail markers.
• Downloaded CSV from the portal’s report page.

Troubleshooting tips

• If the desktop app auto‑updates on launch, wait for completion and rerun the login step.
• If pop‑up blockers interfere with “Export,” allow the site or right‑click and “Save link as…”.

Variations and extensions

• Add step durations and a total runtime metric to the HTML report.
• Take DOM snapshots in the browser by saving page HTML for diffing across runs.

7) Invoice Processing and Accounting Entry

Complete prompt

Automate intake of vendor invoices and entry into accounting software (desktop or web).

Goal:
- From C:\Ops\Downloads\Invoices\[YYYY-MM-DD]\, process all PDFs.
- Extract Vendor, Invoice Number, Invoice Date, Due Date, Currency, Total, and Line Items if available.
- Create draft bills in the accounting system (do not post).
- Prepare a reconciliation CSV and an Outlook draft summary email.

Applications:
- File Explorer
- Edge (web accounting app) or "ContosoBooks" desktop app
- Excel
- Outlook

Steps:
1) For each PDF in the folder:
   - Open the PDF; copy visible fields. If selectable text is poor, use "Save As Text" if available; otherwise, capture key data manually by reading the content.
   - Normalize dates to ISO YYYY-MM-DD and amounts to decimal with currency code.
2) Accounting entry:
   - Open the accounting app.
   - Navigate to Bills/Vendors → New Bill.
   - Enter Vendor (auto-complete), Invoice Number, Dates, and Total.
   - Attach the PDF to the draft bill (Upload or Drag & Drop).
   - Save as Draft, not Posted.
3) Excel reconciliation:
   - Append a row per invoice to C:\Ops\Reports\Invoices_Inbox_[YYYY-MM-DD].xlsx with columns:
     Vendor, InvoiceNumber, InvoiceDate, DueDate, Currency, Total, DraftLink/ID, FilePath, Status.
4) Outlook summary draft:
   - Create a draft email to [email protected] with subject "Invoice Intake – [YYYY-MM-DD]".
   - Body: a table of invoices processed, totals by currency, and notes on missing fields.
   - Attach the reconciliation Excel.
5) Log success/fail per file to C:\Ops\Backups\Logs\Invoices_[YYYY-MM-DD].txt.

Verification:
- Each PDF results in a Draft bill with an attachment present.
- The Excel reconciliation file has a row for each PDF and a nonzero Total.
- Outlook draft exists with the attachment and table.

Constraints:
- Do not post or approve bills automatically.
- If Vendor not found, create a note "Vendor missing – manual setup required" and skip posting the draft.

Error handling:
- If a PDF is password-protected, log "encrypted" and skip.
- If duplicate Invoice Number detected by the system, mark "Duplicate suspected" in Status.

Prerequisites and setup steps

• Folder with sample vendor PDFs named consistently (Vendor_Inv#_Date.pdf helpful but not required).
• Accounting app access with permissions to create drafts.
• Excel trusted location for C:\Ops\Reports\ to allow quick saves.

Expected behavior and verification

• Draft bills created with PDFs attached; no postings made.
• Reconciliation workbook and summary draft email ready for AP review.

Troubleshooting tips

• If OCR is unreliable, zoom PDF to 150–200% for clearer capture.
• If auto‑complete mismatches Vendor names, verify the Vendor ID in the accounting UI before saving.

Variations and extensions

• Tag international invoices for FX review if Currency ≠ local currency.
• Split by department by reading cost center codes from the PDF or email body.

8) Social Media Scheduling and Content Posting

Complete prompt

Automate scheduling of approved social posts to multiple platforms via web UIs.

Goal:
- From C:\Ops\Reports\SocialQueue_[YYYY-MM-DD].xlsx, schedule posts on LinkedIn and X (Twitter) using their web interfaces or a scheduler tool if available.
- Each row: Platform, PostText, ImagePath (optional), TargetTime (local), LinkURL (optional), Tags.
- Create a per-platform results sheet with status and scheduled URL if visible.

Applications:
- Microsoft Edge
- Excel
- File Explorer

Steps:
1) Open Excel file and read rows.
2) For each row:
   - Open the target platform in a new tab:
       * LinkedIn: https://www.linkedin.com/
       * X: https://x.com/
   - If a scheduler tool like Buffer or Hootsuite is preferred and available, use it instead (same credentials).
3) LinkedIn:
   - Click "Start a post".
   - Paste PostText; attach ImagePath if provided; add LinkURL if provided.
   - Choose "Schedule" (clock icon); set TargetTime; confirm.
   - Save scheduled post; capture any confirmation link or text.
4) X:
   - Click "Post" composer; paste PostText; attach ImagePath if provided.
   - Open "Schedule" and set TargetTime; confirm.
   - Save scheduled post; capture confirmation where possible.
5) Update Excel with Status (Scheduled/Failed) and any confirmation link.
6) Save the updated workbook to C:\Ops\Reports\SocialQueue_[YYYY-MM-DD]_RESULTS.xlsx.

Verification:
- Result workbook contains a row-by-row Status.
- Scheduled confirmation UI text captured for at least one post per platform.
- If scheduling not available (policy), create Drafts instead and mark "Drafted".

Constraints:
- Do not change account settings.
- Respect platform limits; introduce 10–20s delay between posts.

Error handling:
- If login is required, pause and request I sign in; do not store credentials.
- If an image path is invalid, mark Failed and continue to next row.

Prerequisites and setup steps

• Platform accounts with permission to schedule posts.
• Images available locally under the specified paths; ensure they meet platform dimensions/size limits.

Expected behavior and verification

• Posts scheduled or drafted per the queue, with a results workbook saved.

Troubleshooting tips

• If the “Schedule” control is hidden behind a three‑dot menu, instruct the agent to look for clock icons or “Schedule” labels.
• If a link preview blocks scheduling, remove the preview and keep the raw URL in the text.

Variations and extensions

• Use a unified scheduler UI if company policy requires a single pane of glass.
• Insert UTM parameters at scheduling time and track them in the results workbook.

9) System Maintenance: Updates, Cleanup, and Monitoring

Complete prompt

Automate weekly system maintenance on Windows 11.

Goal:
- Check for Windows Updates and Microsoft Store app updates.
- Clear temp files safely.
- Collect system health info and export to a report.

Applications:
- Windows Settings (Windows Update)
- Microsoft Store
- Disk Cleanup (cleanmgr) or Storage Sense via Settings
- Windows Terminal (PowerShell)
- Notepad

Steps:
1) Open Settings > Windows Update:
   - Click "Check for updates".
   - If updates are available, download but do NOT restart automatically; wait for my confirmation if a restart is required.
   - Take a screenshot of the update list.
2) Open Microsoft Store:
   - Go to Library > Get updates; update all and screenshot result.
3) Run cleanup:
   - Use Settings > System > Storage > Temporary files; select Downloaded Windows Update cleanup, Temporary files, Thumbnails; do not select Recycle Bin.
   - Alternatively run cleanmgr /LOWDISK and select safe categories; confirm.
4) Health report:
   - In PowerShell, run:
     - systeminfo | Out-File C:\Ops\Backups\Logs\SystemInfo_[YYYY-MM-DD].txt
     - Get-Process | Sort-Object CPU -Descending | Select -First 20 | Out-File C:\Ops\Backups\Logs\TopCPU_[YYYY-MM-DD].txt
     - Get-EventLog -LogName System -Newest 200 | Out-File C:\Ops\Backups\Logs\Events_[YYYY-MM-DD].txt
5) Compose a consolidated maintenance report at C:\Ops\Backups\Logs\Maintenance_[YYYY-MM-DD].txt summarizing:
   - Update status
   - Store updates
   - Cleanup categories and estimated space freed
   - Health metrics files created
6) If a restart is pending, present a prompt: "Restart now?" Wait for my decision.

Verification:
- Logs exist with today's date.
- At least one screenshot captured for updates and Store results.
- Disk cleanup freed > 100 MB or report that nothing to clean.

Constraints:
- Do not restart without consent.
- Do not remove Downloads, Recycle Bin, or personal folders.

Error handling:
- If Windows Update UI errors, run "wuauclt /detectnow" in PowerShell and retry.
- If cleanmgr not found, use Settings Storage Sense flow.

Prerequisites and setup steps

• Administrative rights may be required for some cleanup categories.
• Ensure PowerShell execution is permitted for built‑in cmdlets (default OK).

Expected behavior and verification

• Updates checked, Store apps updated, and logs consolidated without forced restart.

Troubleshooting tips

• If Storage settings take time to calculate, wait until categories populate before selecting.
• If EventLog cmdlet is slow, reduce the number of events fetched.

Variations and extensions

• Export a CSV of installed applications and versions for audit.
• Add Defender quick scan via Windows Security UI and record status.

10) Multi‑Application Workflow: CRM to Spreadsheet Pipeline

Complete prompt

Automate a daily CRM-to-spreadsheet pipeline combining web and desktop apps.

Goal:
- Extract the "Open Opportunities" view from the CRM web portal.
- Normalize data and enrich with a local lookup table.
- Save to an Excel workbook with calculated fields and a dashboard.

Applications:
- Microsoft Edge
- Excel
- File Explorer
- Notepad (for error notes)

Inputs:
- CRM URL: https://crm.contoso.com/opportunities?status=open
- Local lookup: C:\Ops\Reports\OwnerLookup.xlsx (OwnerEmail, Region, Manager)

Steps:
1) Open Edge to the CRM URL. Login if needed (pause for my credentials).
2) Set the view to "Open Opportunities", show 100 rows per page, and include columns:
   - Opportunity Name, Owner, Stage, Amount, Close Date, Account, Probability.
3) Export:
   - If an Export button exists, click and download CSV to C:\Ops\Downloads\CRM_OpenOpps_[YYYY-MM-DD].csv.
   - If not, select the table and copy/paste to Excel, then clean headers and save as CSV.
4) Open Excel and import the CSV to a sheet "Raw".
5) Convert columns to the right types and add calculated fields:
   - WeightedAmount = Amount * Probability
   - DaysToClose = CloseDate - TODAY()
6) Open OwnerLookup.xlsx and VLOOKUP/XLOOKUP OwnerEmail to bring Region and Manager.
7) Build a PivotTable on a "Dashboard" sheet:
   - Rows: Region, Manager
   - Values: Sum of Amount, Sum of WeightedAmount
   - Filters: Stage (multi-select)
   - Add a slicer for Stage and Owner.
8) Insert a stacked column chart for WeightedAmount by Region.
9) Save to C:\Ops\Reports\CRM_OpenOpps_[YYYY-MM-DD].xlsx. Also export the Dashboard sheet to PDF.

Verification:
- Raw sheet row count > 0 and matches the exported CSV rows.
- Lookup fields populate for >= 90% of rows (warn otherwise).
- Dashboard pivot and chart render without #N/A.

Constraints:
- Do not change CRM records.
- Remove PII such as phone numbers before exporting if visible.

Error handling:
- If the export fails due to pop-up blocks, allow pop-ups for the site and retry.
- If headers change, prompt me to confirm new column names before proceeding.

Prerequisites and setup steps

• CRM account with read‑only access to the Opportunities view.
• OwnerLookup.xlsx with accurate OwnerEmail mappings.

Expected behavior and verification

• Clean workbook with enriched data and an actionable dashboard.

Troubleshooting tips

• If CSV delimiter issues arise, set Excel to use comma and re‑import with Data > From Text/CSV with delimiter selection.
• If probability is displayed as a percentage string, strip the “%” and divide by 100 in Power Query or Excel.

Variations and extensions

• Append historical snapshots into a “History” sheet with a RunID timestamp to enable trend charts.
• Publish the dashboard as a PDF to Teams or SharePoint via the desktop UI.

Feature matrix: prompts vs. capabilities

Practical code helpers you can allow in prompts

While these playbook prompts rely mainly on UI actions, adding lightweight scripts as helpers can improve accuracy and speed. You can paste these snippets into your prompts under a section “Helper commands you may run in PowerShell.”

PowerShell: Safe hash and manifest generation

# Generate a manifest for all files under a directory
Param([string]$Root="C:\Ops\Backups\2026-07-17",[string]$Out="C:\Ops\Backups\2026-07-17\manifest.csv")
"Name,FullPath,Size,HashSHA256" | Out-File -FilePath $Out -Encoding utf8
Get-ChildItem -Path $Root -Recurse -File |
  ForEach-Object {
    $h = Get-FileHash -Algorithm SHA256 -Path $_.FullName
    "{0},{1},{2},{3}" -f $_.Name,$_.FullName,$_.Length,$h.Hash
  } | Out-File -FilePath $Out -Append -Encoding utf8

PowerShell: System snapshot

$d = Get-Date -Format "yyyy-MM-dd"
systeminfo | Out-File "C:\Ops\Backups\Logs\SystemInfo_$d.txt"
Get-Process | Sort-Object CPU -Descending | Select-Object -First 20 |
  Format-Table -AutoSize | Out-String | Out-File "C:\Ops\Backups\Logs\TopCPU_$d.txt"
Get-EventLog -LogName System -Newest 200 |
  Format-Table -AutoSize | Out-String | Out-File "C:\Ops\Backups\Logs\Events_$d.txt"

Verification patterns you should encode in prompts

• File existence checks with expected size thresholds and row counts.
• UI label presence checks when entering critical forms (e.g., “Save as Draft” visible).
• Confirmation dialogs before irreversible actions (send, delete, post, restart).
• Timeouts with retries and backoff when waiting on network operations.

Building custom Computer Use workflows

To extend beyond this playbook, structure each new workflow around strong contracts and verifiable gates:

1. Define data inputs and outputs precisely. State file names, columns, formats, and destinations.
2. List allowed applications and web domains to narrow the action space and reduce surprises.
3. Specify UI anchors (exact labels or icons) and fallbacks (menu paths, keyboard shortcuts).
4. Include error branches (“if X not visible, do Y”) to minimize abandonment on minor UI changes.
5. Mandate evidence capture (screenshots, logs) and summarize success criteria at the end.

For consistency across teams, standardize directory layouts, filename conventions, and a shared glossary of category names and labels. Where possible, have prompts reference the same core terms your users see in the UI to reduce ambiguity.

Combining prompts for complex automations

Many real‑world processes chain tasks across apps and days. Combine prompts by:

• Declaring upstream artifacts: “Assume the output of Prompt 2 exists at path … and must be refreshed first.”
• Splicing steps: take the email drafts from Prompt 1 and attach PDFs created in Prompt 7 before sending.
• Scheduling: run the file backup (Prompt 4) after the maintenance sweep (Prompt 9) to archive logs and reports together.

When chaining, re‑state key verification steps at each handoff to keep errors from propagating. For example, confirm “Row count >= N and column headers match expected set” before using an imported CSV downstream.

Performance optimization tips

• Reduce visual clutter: close nonessential windows, switch to Light theme where labels are higher contrast, and set display scaling to 100–125%.
• Prefer direct exports over copy/paste when available; CSV/Excel exports preserve structure and reduce OCR dependence.
• Cache credentials with Windows Hello or single‑sign‑on for faster, safer logins (still require consent to proceed).
• Batch similar steps: e.g., gather all PDFs first, then switch to the accounting app to create drafts in sequence.
• Cap per‑run scope (e.g., 50 emails) to limit drift and review time; run multiple batches if needed.

Security best practices

• Never embed passwords or API keys in prompts. Use OS‑mediated credential prompts and keep the human in the loop for consent.
• Store logs and screenshots in a secure directory with appropriate ACLs; redact PII when sharing.
• Use dedicated test accounts for development and staging. Promote prompts to production only after review.
• Whitelist allowed apps and domains per prompt; explicitly prohibit sensitive actions like deleting or mass mailing unless confirmed.
• Document a rollback path for each automation (e.g., drafts only, restore from backup, review queue).

Frequently asked questions

How do I make these prompts robust to minor UI changes?

Use multiple anchors: label text, icon names, and menu paths. Provide fallbacks (“If the Schedule button is not visible, click the three‑dot menu and select Schedule”). Keep selectors abstract (“button labeled ‘Export’”) but constrain the window title and section you expect to be in. Include verification after each navigation hop.

Can Computer Use run headless or on a server?

Computer Use relies on visual context, so it needs a visible or virtual display surface. You can run it on a VM with a GPU‑backed session or a virtual desktop infrastructure, but ensure a consistent resolution and avoid screen locks during runs. For unattended tasks, keep the session active and restricted to the automation user.

What about localization—will these prompts work on non‑English UIs?

They rely on specific English labels. For localized environments, translate labels and menu paths in the prompt. Where possible, target icons, positions, or shortcuts that are language‑agnostic, but do not rely solely on spatial position because layouts can vary.

How do I review and approve before actions like sending emails or posting?

Include an explicit confirmation gate: “Present a final confirmation dialog … wait for my input.” Require human consent before sending, posting, or deleting. Draft‑only patterns (Outlook drafts, scheduler drafts) provide safe review buffers.

What logging should I require for auditability?

At a minimum: timestamped step logs, file manifests for outputs, and screenshots for key decisions. Save logs under C:\Ops\Backups\Logs\ with a consistent naming scheme. For sensitive data, store logs on a restricted share with retention policies.

How do I detect and handle duplicates (e.g., invoices)?

Leverage app‑level duplicate checks (error banners) and add your own: keep a local CSV of processed identifiers (Invoice Number + Vendor + Date). Before creating a new record, search the app UI for the identifier; if found, mark “Duplicate suspected” and skip.

Is Power Automate Desktop still useful if I use Computer Use?

Yes. Power Automate Desktop gives deterministic flows with strong control constructs and deep Windows actions. Computer Use is ideal when UI changes, when apps lack connectors, or for rapid prototyping. Many teams use both: codify stable, high‑volume tasks in PAD and keep flexible workflows in Computer Use. See

Understanding how these capabilities compare across different AI platforms provides valuable context for tool selection. Our head-to-head comparison in Claude Opus 4.7 vs OpenAI Codex for Indie Shipping: Which Should You Choose in 2026? evaluates the strengths and trade-offs of competing approaches.

for a detailed comparison.

How do I migrate a working prompt into a shared, versioned asset?

Store the prompt text in source control with a README documenting prerequisites, environment assumptions, and expected outputs. Tag releases tied to application versions (e.g., “Outlook 2409 prompt”). Maintain a changelog noting UI label changes and newly added verification steps.

Conclusion

These 10 prompts operationalize Codex Computer Use for real Windows work: triaging Outlook, shaping Excel data, researching on the web, backing up files, extracting meeting actions, testing UIs, processing invoices, scheduling social content, maintaining systems, and bridging CRM to spreadsheets. By encoding explicit verification steps, consent gates, and logging, you can reach production‑grade reliability while maintaining the agility that makes Computer Use compelling. Start with one prompt, harden it with evidence and constraints, and then compose multi‑step runbooks that move business outcomes end‑to‑end without waiting on new APIs or integrations.