Override the AI model for specific actions in your workflow prompts
The model parameter lets you specify which AI model to use for individual focused_action and screenshot actions that include extract_prompt, directly in your workflow prompt.
For comprehensive documentation on model configuration, including workflow-level settings, custom configurations, and supported providers, see Model Configuration.
Use focused_action with model="Sonnet 4 (Thinking)" to verify the order statusTake screenshot with extract_prompt="Extract invoice data" and model="Sonnet 4.5"
Tab autocomplete: Type model → press Tab → inserts model=""
Direct typing: Type model="" manually
With your cursor inside the quotes, a dropdown appears showing all available models. Hover to see model details including computer use support.The dropdown also includes System Default, which clears the override and lets Cyberdesk choose the normal default for that action.
For focused_action: Prefer models marked as “computer use” models. The editor warns if you pick another configured model, because non-computer-use models may not handle clicking, typing, and navigation reliably.For extract_prompt: Any vision-capable model works since it only reads the screen.
Navigate to the patient records system and log in.Use focused_action with model="Sonnet 4 (Thinking)" to find the patient whose name is {patient_name} and MRN ends with {mrn_suffix}. This requires careful reasoning to match the correct record.Once on the patient chart, take a screenshot with extract_prompt="Extract all vital signs as JSON: {blood_pressure, heart_rate, temperature, oxygen_saturation}" and model="Sonnet 4.5"Navigate to the lab results tab and extract data from each page:- Take screenshot with extract_prompt="Extract lab results as JSON array" and model="Sonnet 4.5" and process_async="batch"- Scroll down- Repeat until all results captured
Open the invoice management portal.Search for invoice {invoice_number} and open the details page.Use focused_action with model="Sonnet 4 (Thinking)" to verify this is the correct invoice by checking the customer name matches {customer_name} and the date is {invoice_date}.Take a screenshot with extract_prompt="Extract complete invoice data: invoice_number, date, customer info, all line items with descriptions and amounts, subtotal, tax, and total. Format as detailed JSON." and model="Sonnet 4.5" and process_async="run"Continue to the next invoice while extraction runs in background.
Navigate to the appointment scheduling system.Use focused_action with model="Sonnet 4 (Thinking)" to analyze the calendar and find an available slot that meets ALL criteria:- Day is {preferred_day}- Time is between {earliest_time} and {latest_time}- Provider is {preferred_doctor}- Duration allows for {appointment_length} minutesIf no slot meets all criteria, use focused_action with model="Sonnet 4 (Thinking)" to find the closest alternative and explain the tradeoff.Click to book the selected slot.
Open the product catalog and navigate to category {category_name}.For each page of products:- Take screenshot with extract_prompt="Extract all visible products as JSON: {sku, name, price, stock_status, rating}" and model="Sonnet 4.5" and process_async="batch"- Scroll to next page- Repeat until end of catalogAll extractions use the same fast model for consistent results and cost.
Submit the insurance claim form.Use focused_action with model="Sonnet 4 (Thinking)" to analyze the confirmation screen:- Verify the claim ID is displayed- Check that all submitted fields match our input values- Identify any warnings or additional steps required- Save the claim ID as {{claim_id}}If any discrepancies are found, use declare_task_failed with details.Take screenshot with extract_prompt="Extract confirmation details: claim_id, submission_date, estimated_processing_time, any reference numbers" and model="Sonnet 4.5"
Open the monthly financial report.Fast clipboard extraction for key fields:- Triple-click Report Date field, use copy_to_clipboard with key "report_date"- Triple-click Total Revenue field, use copy_to_clipboard with key "total_revenue"Vision extraction for complex content:- Take screenshot of the revenue breakdown chart with extract_prompt= "Extract revenue by category from this chart as JSON" and model="Sonnet 4.5"- Take screenshot of the expenses table with extract_prompt= "Extract all expense line items as JSON array" and model="Sonnet 4.5" and process_async="batch"Dynamic verification:- Use focused_action with model="Sonnet 4 (Thinking)" to verify the totals in the summary section match the detailed breakdowns
