Goal

Generate Salesforce Flow metadata by running the required 3-step MCP pipeline (fetchGroundedObjectMetadata → flowElementSelection → flowElementGeneration) and return the flow XML.

When to Use This Skill

Use this skill when you need to:

• Create any type of Flow (Screen, Autolaunched, Record-Triggered, Scheduled)
• Generate Flow metadata XML
• Automate business processes without code
• Build user-guided workflows or background automation
• Troubleshoot deployment errors related to Flows

Specification

Flow Metadata Specification

Overview

Salesforce Flows are powerful automation tools that enable complex business process automation without code. Flows can collect and process data through interactive screens, execute logic and calculations, manipulate records, call external services, and trigger based on various events. Flow types include Screen Flows (user-guided), Autolaunched Flows (background processing), Record-Triggered Flows (database events) and Scheduled Flows (time-based).

Flow Generation Pipeline

MANDATORY: You MUST follow this exact 3-step pipeline. No exceptions. No shortcuts. No skipping steps. Do NOT manually create flow metadata XML or attempt to generate flow metadata outside of this pipeline. Do NOT attempt to use any other tool, API, or method to generate flow metadata. This pipeline is the ONLY supported way to generate flows. Any deviation will produce invalid or broken metadata.

MCP Connection Details

All 3 pipeline steps MUST be called using this MCP tool:

• MCP Tool Name: execute_metadata_action
• The action parameter selects which pipeline step to run: "fetchGroundedObjectMetadata", "flowElementSelection", or "flowElementGeneration"

Flow generation is a strict 3-step pipeline. ALL steps must be called in order. Every step is required. There is no alternative approach — this is the only way to generate flow metadata:

Step 1 (REQUIRED): Fetch Grounded Object Metadata (fetchGroundedObjectMetadata)

Fetches org schema metadata relevant to the flow generation request. This step is mandatory and must always be called first.

Inputs (all required):

• userPrompt (STRING, REQUIRED): The user's natural language request
• inflightMetadata (ARRAY, REQUIRED): Custom objects/fields from local sfdx project. Use empty array [] if none needed.

Outputs:

• groundingMetadata (STRING): Grounded object metadata for org schema relevant to the request, returned as a JSON string. You must pass this directly to Step 2 — it is already a string and does not need to be serialized again.

Step 2 (REQUIRED): Flow Element Selection (flowElementSelection)

Selects flow elements (assignments, decisions, record ops, etc.) and their connections based on the user prompt and grounded metadata. This step is mandatory and must be called after Step 1.

Inputs (all required):

• userPrompt (STRING, REQUIRED): The user's natural language request (must be the same value as Step 1)
• groundingMetadata (STRING, REQUIRED): Org schema metadata (must be the exact string returned from Step 1 output — pass it directly, do NOT serialize it again)
• operationId (STRING, REQUIRED): Operation ID (use empty string "" for first call)

Outputs:

• operationId (STRING): Operation ID. You must pass this to Step 3.
• userOutput (STRING): Reasoning for next steps. You can show this to the user.

Step 3 (REQUIRED): Flow Element Generation (flowElementGeneration)

Generates flow metadata element by element. This step is mandatory and must be called after Step 2. Must be called repeatedly in a loop until isComplete is true.

Inputs (all required):

• operationId (STRING, REQUIRED): Operation ID from Step 2 output
• requestSource (STRING, REQUIRED): The source of the request. Use "A4V" to get flow metadata in XML format.

Outputs:

• isComplete (BOOLEAN): Indicates if the flow generation is complete. You must check this value.
• result (STRING): Result of the flow element generation. Contains the final flow metadata only when isComplete is true.

MANDATORY: Loop until complete. NEVER pause or ask the user to confirm continuation.

• A flow can have any number of elements (10, 15, or more). Each call generates one element at a time, so you may need many iterations. This is expected and normal.
• Call flowElementGeneration with the operationId from Step 2 and requestSource (use "A4V" for XML output, empty string or other value for JSON).
• Check the isComplete output and the result field after each call.
• If isComplete is false and no errors are returned, you MUST call flowElementGeneration again with the same operationId from Step 2. Do NOT ask the user if they want to continue. Do NOT pause. Do NOT summarize progress mid-loop. Just keep calling.
• Do NOT stop until isComplete is true or the invocable action returns errors. There is no maximum number of iterations — keep going regardless of how many calls it takes.
• When isComplete is true, extract the flow metadata from the result field.
• If errors are returned, stop the loop and surface the error to the user.

STRICT CONSTRAINTS (CRITICAL) — These rules apply to the XML returned by the generation pipeline:

• DO NOT modify the content, values, or child nodes inside any block.
• DO NOT add new nodes, tags, attributes, or text (do not add missing labels, X/Y coordinates, etc.).
• DO NOT remove any existing nodes.

inflightMetadata Format

DATA TYPE: ARRAY (not string)

STRICT NAMING CONVENTION - MUST FOLLOW EXACTLY:

When custom objects are needed (sample format showing multiple field data types):

[
  {
    "type": "CustomObject",
    "apiName": "CustomerRequest__c",
    "label": "Customer Request",
    "fields": [
      {
        "apiName": "Status__c",
        "type": "Picklist",
        "label": "Status",
        "values": ["New", "In Progress", "Completed"]
      },
      {
        "apiName": "Priority__c",
        "type": "Number",
        "label": "Priority"
      },
      {
        "apiName": "AssignedTo__c",
        "type": "Lookup",
        "label": "Assigned To",
        "referenceTo": "User"
      },
      {
        "apiName": "Description__c",
        "type": "Textarea",
        "label": "Description"
      },
      {
        "apiName": "Email__c",
        "type": "Email",
        "label": "Contact Email"
      },
      {
        "apiName": "DueDate__c",
        "type": "Date",
        "label": "Due Date"
      },
      {
        "apiName": "IsUrgent__c",
        "type": "Boolean",
        "label": "Is Urgent"
      },
      {
        "apiName": "Amount__c",
        "type": "Currency",
        "label": "Amount"
      }
    ],
    "relationships": []
  }
]

Supported field types: Text, Textarea, Number, Picklist, Lookup, Email, Phone, URL, Date, Datetime, Boolean, Checkbox, Currency, Percent

When no custom objects needed:

[]

MANDATORY Decision Logic for inflightMetadata (DATA TYPE: ARRAY)

1. REQUIRED - First: Scan the local sfdx project for custom objects and fields that are relevant to the user's flow request.
2. If relevant custom objects ARE found: You MUST extract and pass them as an array of structured objects (see format above)
3. If NO relevant custom objects found: You MUST pass an empty array [] (NOT the string "[]")
4. NEVER: Pass text descriptions, instructions, or string representations in inflightMetadata
5. MANDATORY: The data type MUST be ARRAY, not STRING

Instructions for Vibes when custom objects ARE relevant:

• Extract the object metadata and map to JSON properties:
  • apiName: The object's API name (with __c suffix for custom objects)
  • label: The object's display label
  • type: Set to "CustomObject"
  • fields: Array of field objects, each containing:
    • apiName: The field's API name (with __c suffix for custom fields)
    • type: The field type (Text, Number, Picklist, Lookup, etc.)
    • label: The field's display label
    • values: (Picklist only) Array of picklist values
    • referenceTo: (Lookup only) The target object API name
• Include only objects and fields that are relevant to the flow being generated

Mandatory Enhancement Rules

• userPrompt: REQUIRED.
  • If the user requests a single flow: use the user's prompt as-is.
  • If the user requests multiple flows: you MUST split the request and write a separate, focused userPrompt for each individual flow. Each userPrompt must describe only ONE flow. Do NOT pass the entire multi-flow request as a single userPrompt. See the multiple flows section below for examples.
• inflightMetadata: REQUIRED. Always use ARRAY data type.
  • MUST use [] (empty array) when no custom objects needed
  • MUST use structured array of objects when custom objects are relevant
  • NEVER use string "[]" - this is incorrect
  • NEVER use text descriptions - only structured object metadata

Flow Pattern Guidance

The pipeline selects flow elements from your userPrompt. Vague prompts lead the pipeline to pick wrong structures, producing metadata that fails deployment. When the request matches one of the patterns below, make the intent explicit in the userPrompt you pass to Step 1 and Step 2 so the pipeline selects the correct elements.

Scheduled flows ("daily", "weekly", "every Sunday", "runs once a week at 10PM")

A recurring, time-based flow is a Scheduled flow. The schedule lives on the start element: triggerType is Scheduled and a <schedule> block holds <frequency> (e.g. Weekly), <startDate>, and <startTime>. Do NOT expect startDate/startTime on a FlowScheduledPath element — a scheduled flow's cadence is on <start><schedule>, not a scheduled path.

To select records for a scheduled flow, put the record criteria in the start element's filters (with <object> and <filters> on <start>), which runs the flow once per matching record with $Record bound to each. Prefer this over adding a Loop that re-queries and iterates the same object — a start-filtered scheduled flow does not need a Loop to walk the triggering object's records.

When the prompt says "runs once a week / every Sunday / daily at a time", the userPrompt should state: scheduled trigger, the frequency, the start time, and the record filter on the triggering object.

Counting related records ("count all related X", "number of X")

To store a count of records, use a single assignment element with <operator>AssignCount</operator>, assigning from the collection (the record-lookup result) into a Number variable. Do NOT emit two assignment elements for one count, and do NOT give two assignment elements the same name — duplicate assignment names, or two assignments doing one logical count, fail deployment. One lookup → one AssignCount assignment → one record update.

Do not invent actions the prompt did not request

Generate only the elements the prompt asks for. If a prompt names a trigger but not an action (e.g. "create a flow for when a record is created" with no stated behavior), do NOT add a Chatter post, email, or other action that was not requested. An unrequested action such as a chatterPost produces metadata that references undefined types and fails deployment. When the requested behavior is genuinely absent, generate the minimal valid trigger without inventing side effects.

MANDATORY: Multiple Flows = Multiple Separate Pipelines

FIRST: Before calling any pipeline step, check if the user's request contains multiple flows. If it does, you MUST split it into separate single-flow prompts. Each flow gets its own 3-step pipeline with its own userPrompt that describes ONLY that one flow.

NEVER pass a multi-flow request as a single userPrompt field. NEVER club multiple flow descriptions into one userPrompt.

When the user requests multiple flows (e.g., "Create flows for my app: 1) ... 2) ... 3) ..."), you MUST:

1. Split the request into separate individual flow descriptions.
2. Run a separate 3-step pipeline for each flow, using a userPrompt that describes ONLY that one flow.
3. Execute ALL pipelines SEQUENTIALLY — one after another, NEVER in parallel. Do NOT stop after the first flow. Do NOT wait for the user to ask you to continue. Do NOT summarize and stop. Keep going until every requested flow has been fully generated.

WRONG - Multiple flows clubbed into one userPrompt:

{
  "userPrompt": "Create flows for the app: 1) Record-Triggered Flow on ResourceAllocation__c to update Resource__c. 2) Screen Flow to allocate resources. 3) Record-Triggered Flow on Supply__c to auto-flag Low_Stock__c.",
  ...
}

CORRECT - Separate call for EACH flow:

Flow 1 - Step 1 (fetchGroundedObjectMetadata):

{
  "userPrompt": "Create a Screen Flow named Tenant_Onboarding that captures tenant details, selects a Unit__c with Status__c = 'Vacant', creates Lease__c...",
  "inflightMetadata": [...]
}

Then call Step 2 (flowElementSelection) with the groundingMetadata from Step 1, then Step 3 (flowElementGeneration) with the operationId from Step 2.

Flow 2 - Step 1 (fetchGroundedObjectMetadata):

{
  "userPrompt": "Create an Autolaunched Flow named Generate_Onboarding_Checklist that given a Lease__c Id input, queries OnboardingTask__c...",
  "inflightMetadata": [...]
}

Then call Step 2 and Step 3 for this flow.

Flow 3 - Step 1 (fetchGroundedObjectMetadata):

{
  "userPrompt": "Create a Record-Triggered Flow named Sync_Unit_On_Lease_Changes that on insert and update of Lease__c...",
  "inflightMetadata": [...]
}

Then call Step 2 and Step 3 for this flow.

Mandatory Rules:

• If there are N flows to generate, there MUST be N separate 3-step pipelines and ALL N pipelines MUST be executed. No exceptions. Do NOT stop after generating only one flow.
• You MUST fully complete the current flow's 3-step pipeline (including looping Step 3 until isComplete is true or errors are returned) BEFORE starting the next flow's pipeline. Do NOT interleave or parallelize pipelines across flows. Everything is SEQUENTIAL — NEVER parallel.
• After completing a flow's pipeline, immediately start the next flow's pipeline. Do NOT pause, summarize, or wait for user confirmation between flows.
• For each flow, you MUST scan the local sfdx project to populate inflightMetadata with custom objects/fields specific to that flow prompt.
• Each flow pipeline MUST have its own inflightMetadata containing only the objects/fields relevant to that particular flow.

Example Tool Calls

Example 1: Standard objects only (no custom objects)

Step 1 - fetchGroundedObjectMetadata:

{
  "userPrompt": "Create a scheduled-triggered Flow named Daily_Good_Morning that runs daily at 6:00 AM and sends an email to the running user saying good morning.",
  "inflightMetadata": []
}

Step 2 - flowElementSelection:

{
  "userPrompt": "Create a scheduled-triggered Flow named Daily_Good_Morning that runs daily at 6:00 AM and sends an email to the running user saying good morning.",
  "groundingMetadata": "<groundingMetadata string from Step 1 — pass directly, do not serialize again>",
  "operationId": ""
}

Step 3 - flowElementGeneration (call in a loop):

{
  "operationId": "<operationId from Step 2>",
  "requestSource": "A4V"
}

Call repeatedly with the same operationId until isComplete is true or errors are returned. A flow can have any number of elements, so expect multiple iterations. When isComplete is true, extract the flow metadata from the result field. Use "requestSource": "A4V" to get flow metadata in XML format.

Example 2: With custom objects from local sfdx project

Step 1 - fetchGroundedObjectMetadata:

{
  "userPrompt": "Create a flow that updates the status of a Customer Request when it's assigned",
  "inflightMetadata": [
    {
      "type": "CustomObject",
      "apiName": "CustomerRequest__c",
      "label": "Customer Request",
      "fields": [
        {
          "apiName": "Status__c",
          "type": "Picklist",
          "label": "Status",
          "values": ["New", "In Progress", "Completed"]
        },
        {
          "apiName": "AssignedTo__c",
          "type": "Lookup",
          "label": "Assigned To",
          "referenceTo": "User"
        }
      ],
      "relationships": []
    }
  ]
}

Step 2 - flowElementSelection:

{
  "userPrompt": "Create a flow that updates the status of a Customer Request when it's assigned",
  "groundingMetadata": "<groundingMetadata string from Step 1 — pass directly, do not serialize again>",
  "operationId": ""
}

Step 3 - flowElementGeneration (call in a loop):

{
  "operationId": "<operationId from Step 2>",
  "requestSource": "A4V"
}

Call repeatedly with the same operationId until isComplete is true or errors are returned. A flow can have any number of elements, so expect multiple iterations. When isComplete is true, extract the flow metadata from the result field. Use "requestSource": "A4V" to get flow metadata in XML format.

CRITICAL Verification Checklist (MUST VERIFY BEFORE AND AFTER EVERY FLOW GENERATION)

Failure to follow this checklist exactly will result in broken or missing flow metadata.

• Pipeline: ALL 3 steps are called in strict order (fetchGroundedObjectMetadata → flowElementSelection → flowElementGeneration). No step is skipped, combined, or replaced by another tool/API. This pipeline is the ONLY way to generate flow metadata.
• No manual metadata: Flow metadata is NOT created, modified, or generated outside the pipeline, and no elements/attributes were added to the returned XML (no <label>, <description>, or any other node). The final XML must be identical to what the pipeline returned. Exception: if the user explicitly requested fixes to validation/deployment errors in already-generated XML, targeted manual edits are permitted.
• userPrompt contains a single flow prompt (multi-flow requests are split, one userPrompt each), holds the flow requirements (NOT inflightMetadata), and is passed identically to Step 1 and Step 2.
• inflightMetadata is ARRAY data type (NOT the string "[]"), is [] when no custom objects are needed, and otherwise holds structured object/field metadata scanned from the local sfdx project — never text descriptions.
• groundingMetadata from Step 1 output is passed directly to Step 2 input (already a string — do NOT serialize it again).
• Step 3 is called in a loop with the same operationId from Step 2, requestSource always "A4V", until isComplete is true or errors are returned — no pausing, no asking the user to continue, no matter how many iterations. Extract the XML from result only when isComplete is true.
• Multi-flow: Each flow's full pipeline is completed before starting the next, SEQUENTIALLY (never in parallel, never interleaved), and ALL requested flows are generated — do NOT stop after the first.