Workflow preview

What the agent will follow (tools, prompts, and workflow steps).

1. Apply the following prompt: Goal:
Understand what the user wants to automate and gather enough detail to design a workflow

Inputs:
The user's description of what they want to automate

Outputs:
A clear problem statement including: what repetitive task is being automated, what inputs the user provides, what the expected output is, which external services or data sources are involved, and where human judgment is required vs where automation can handle it

Constraints:
Ask clarifying questions if the user's request is vague. Identify the key services involved (email, CRM, spreadsheets, OCR, etc.) so you know what tools to search for. Do not start building until you have a clear picture of the full pipeline. Confirm your understanding with the user before proceeding.

Success criteria:
The user has confirmed the automation scope and you have a clear list of capabilities needed for each step

2. Call tool: AgentPMT Workflow Creator (Load Workflow Builder Reference).

Instructions:
Load the full workflow creation reference guide including node types, edge structure, validation rules, design principles, and examples. Internalize this before designing any workflow. Key concepts to retain: workflows are DAGs with tool, prompt, for_each, and notify_human nodes connected by edges. Tool nodes need real product_id values from fetch_tools. Prompt nodes should use structured mode with goal/inputs/outputs/constraints. Graph must be acyclic with unique node IDs.

Parameters:
{"action": "get_instructions"}

3. Call tool: AgentPMT Workflow Creator (Search Tool Catalog).

Instructions:
Search the AgentPMT tool catalog for every capability the workflow needs. Search multiple times with different keywords to cover all aspects. For example, if the user wants to 'process receipts and log expenses', search for 'OCR document extraction', 'file storage upload', 'Google Sheets spreadsheet', etc. Record each tool's _id (use as product_id) and name (use as product_name). If no tool exists for a step, that step becomes a prompt node for AI reasoning or a notify_human node for manual input.

Parameters:
{"action": "fetch_tools"}

4. Apply the following prompt: Goal:
Analyze discovered tools, select the best one for each workflow step, and plan the overall graph architecture

Inputs:
The tool search results from the previous step and the user's automation requirements

Outputs:
A tool selection map (which tool for each step, with product_id and product_name) and a high-level architecture plan showing the node sequence, any for_each loops needed, any branching logic, and where prompt nodes are needed for data transformation between tools

Constraints:
Choose the most specific tool for each task. If multiple tools could work, pick the one whose description best matches the step's purpose. Identify where prompt nodes are needed between tool nodes for: data transformation, decision making, synthesizing information from multiple sources, or applying business logic. Plan for_each nodes when the workflow needs to iterate over a collection (e.g., process each receipt, reconcile each transaction). Only use notify_human when human judgment is genuinely required.

Success criteria:
Every step in the workflow has either a tool assignment or is designated as a prompt/for_each/notify_human node, and the overall flow is clear

5. Apply the following prompt: Goal:
Design the complete workflow graph with all nodes, edges, and configuration ready for creation

Inputs:
The tool selection map and architecture plan from the previous step

Outputs:
A complete workflow specification: all nodes (with IDs, types, labels, tool config or prompt config), all edges (with IDs, from, to, and sourceHandle/targetHandle for for_each nodes), the workflow name, description, time_saved_minutes, and industry_tags

Constraints:
Follow all DAG rules: no cycles, unique node IDs, all edges reference existing nodes, every non-terminal node has an outgoing edge. Use descriptive kebab-case node IDs like 'fetch-data', 'analyze-results'. Use structured mode for all prompt nodes with clear goal/inputs/outputs/constraints. Tool node instructions should explain WHY the tool is called and HOW its output feeds into the next step. Write an SEO-friendly description with keywords for discoverability. Estimate time_saved_minutes honestly based on manual equivalent. Select 2-5 relevant industry tags. Start with visibility private.

Success criteria:
The graph passes all validation rules and every node has complete configuration

6. Call tool: AgentPMT Workflow Creator (Create Workflow).

Instructions:
Create the workflow using the designed graph. Pass the full specification: name, description, time_saved_minutes, visibility (start with private), industry_tags, nodes array, and edges array. Double-check that every tool node has a real product_id from the fetch_tools results. Every workflow description must be unique and keyword-rich. Do not use emojis in names, descriptions, labels, or instructions.

Parameters:
{"action": "create_new"}

7. Call tool: AgentPMT Workflow Creator (Verify Created Workflow).

Instructions:
Fetch the created workflow by skill_id to verify the full graph. Review every node: correct type, correct product_id for tool nodes, clear prompt configuration for prompt nodes, proper for_each setup with parentId on child nodes. Review every edge: correct from/to, sourceHandle/targetHandle for for_each edges. Check that the graph is complete and no nodes are orphaned.

Parameters:
{"action": "fetch_existing"}

8. Apply the following prompt: Goal:
Compare the verified workflow against the original design and fix any discrepancies

Inputs:
The fetched workflow structure from verify-workflow and the original design specification

Outputs:
Either confirmation that the workflow matches the design, or a list of issues found and fixes applied via update_existing

Constraints:
Check: all nodes present with correct types and configuration, all edges connecting the right nodes, no orphaned nodes, tool product_ids match what was discovered, prompt nodes have complete structured config, for_each child nodes have correct parentId. If any issues are found, call AgentPMT-Workflow-Creator with action update_existing to fix them. Then re-fetch to confirm fixes. Do not proceed to publishing until the workflow is verified correct.

Success criteria:
The workflow structure matches the designed specification with zero discrepancies

9. Call tool: AgentPMT Workflow Creator (Publish Workflow).

Instructions:
Publish the verified workflow. First publish sets version to 1.0.0. If the user requested public visibility, update the workflow visibility to public before publishing using update_existing. Publishing creates a frozen snapshot; the live workflow can still be edited afterward.

Parameters:
{"action": "publish", "version_bump": "auto"}

10. Apply the following prompt: Goal:
Present the completed workflow to the user with a clear summary and next steps

Inputs:
The published workflow skill_id, version, and all workflow metadata

Outputs:
A user-friendly summary covering: workflow name and slug, what it automates, number of steps and which tools it uses, industry tags, version number, and how to run it

Constraints:
Include the skill_id so the user can reference it. Explain how to run the workflow using AgentPMT-Workflow-Skills with start_workflow. Mention that they can remix it, update it, or add showcase examples. If the workflow is public, note that others can discover and use it in the marketplace.

Success criteria:
The user understands what was built, how to use it, and how to modify it later