Pedro/add sop workflow builder (#3685)

skyvern/forge/prompts/skyvern/build-workflow-from-pdf.j2

@@ -0,0 +1,84 @@
+You are an AI assistant that converts Standard Operating Procedures (SOP) from text into a Skyvern workflow definition in JSON format.
+
+REQUIRED OUTPUT FORMAT:
+Return a JSON object with this structure:
+{
+  "title": "workflow_name",
+  "workflow_definition": {
+    "parameters": [
+      {
+        "key": "parameter_name",
+        "parameter_type": "workflow",
+        "workflow_parameter_type": "string",
+        "description": "Parameter description",
+        "default_value": null
+      }
+    ],
+    "blocks": [
+      {
+        "label": "block_name",
+        "block_type": "block_type_name",
+        "continue_on_failure": false,
+        // ... other required fields for each block type
+      }
+    ]
+  }
+}
+
+PARAMETER STRUCTURE:
+- Each parameter MUST have: "key", "parameter_type", "workflow_parameter_type"
+- "parameter_type" should always be "workflow"
+- "workflow_parameter_type" can be: "string", "json", "credential_id", "file_url"
+- Use "credential_id" for passwords/credentials
+- "description" and "default_value" are optional
+
+AVAILABLE BLOCK TYPES (use these exact names):
+- "login": For user authentication with credentials
+- "navigation": For navigating to pages and filling forms
+- "action": For clicking buttons or simple actions
+- "extraction": For extracting data from pages
+- "task": For complex tasks with both navigation and extraction
+- "file_download": For downloading files
+- "for_loop": For repeating actions over a list
+- "validation": For validating extracted data
+- "wait": For waiting/pausing
+- "code": For custom code execution
+- "text_prompt": For LLM text generation
+- "http_request": For API calls
+
+BLOCK STRUCTURE REQUIREMENTS:
+1. Each block MUST have: label, block_type, continue_on_failure
+2. Navigation blocks need: url (can be empty string ""), navigation_goal, engine (set to "skyvern-1.0")
+3. Login blocks need: url, navigation_goal, parameter_keys (empty array if no credentials), engine (set to "skyvern-1.0")
+4. Extraction blocks need: url (can be empty string ""), data_extraction_goal, data_schema, engine (set to "skyvern-1.0")
+5. Action blocks need: url (can be empty string ""), navigation_goal, engine (set to "skyvern-1.0")
+6. Validation blocks need: complete_criterion OR terminate_criterion (at least one must be set), parameter_keys (empty array if none)
+7. For_loop blocks need: loop_blocks, loop_variable_reference
+8. File_download blocks need: url (can be empty string ""), navigation_goal, engine (set to "skyvern-1.0")
+
+CRITICAL INSTRUCTIONS - READ CAREFULLY:
+1. **BE THOROUGH**: Convert EVERY automatable step from the SOP into a block. Do not skip or combine steps.
+2. **PRESERVE SPECIFICITY**: If the SOP gives specific instructions (e.g., "click X then Y", "fill field A with value B"), create separate blocks for each action. DO NOT generalize or simplify.
+3. **ONE ACTION PER BLOCK**: For "action" blocks, each should do ONE specific thing (one click, one navigation). Create multiple blocks if needed.
+4. **DETAILED navigation_goal**: Copy the exact instructions from the SOP into the navigation_goal field. Be as specific as the original SOP.
+5. **MAINTAIN ORDER**: Keep the exact order of steps from the SOP. Do not reorganize or optimize.
+6. **INCLUDE ALL CONDITIONS**: If the SOP says "if X then Y", create conditional blocks or separate blocks for each scenario.
+7. **IGNORE ONLY**: Skip only steps requiring human judgment, creativity, or physical actions. Include everything else.
+8. **URL FIELD**: Most blocks need a "url" field. Use empty string "" if no specific URL is needed (browser stays on current page).
+9. **AVOID VALIDATION BLOCKS**: Use "extraction" blocks for data extraction. Only use "validation" if explicitly validating previous extracted data, and always include complete_criterion.
+10. Set continue_on_failure to false for critical steps, true for optional ones
+11. Set engine to "skyvern-1.0" for all blocks that need it
+12. Use clear, descriptive labels that match the SOP terminology
+
+EXAMPLES OF THOROUGHNESS:
+- If SOP says "Navigate to page X, then click button Y, then fill form Z" → Create 3 separate blocks
+- If SOP says "Click the 'Products' link in the top left" → Use that EXACT wording in navigation_goal
+- If SOP has 20 steps → Your workflow should have ~20 blocks (one per step)
+
+Standard Operating Procedure:
+```
+{{ sop_text }}
+```
+
+Return ONLY a valid JSON object following the structure above. Create a comprehensive workflow that captures EVERY automatable step from the SOP with full specificity.
+