Jinba workflow automator

Workflow

When this skill is invoked, the user provides a plain-English description of an enterprise automation they want to build. Your job is to parse that description and emit a complete, structured JSON automation definition that a developer can import into a real execution engine.

This skill produces the definition only. Runtime execution, credential provisioning, and deployment to an engine like n8n, Temporal, or Zapier are out of scope for v1 — that is the productionization gap the developer bridges after receiving this scaffold.

Step 1 — Parse the natural-language workflow description

Read the user's description carefully and extract every distinct component:

• Trigger: What event starts the workflow? Identify the source system (e.g., Google Sheets row added, webhook POST, cron schedule, form submission) and the specific event type.
• Steps: List every distinct action in order. Each sentence or clause describing an action is typically one step.
• Conditions / branching: Identify any "if X then Y" logic. Note the condition expression and which steps it gates.
• Tools and integrations: For each step, identify the third-party service or API it requires (e.g., Clearbit for enrichment, HubSpot for CRM, Slack for notification, SendGrid for email).
• Data flow: Note which outputs from earlier steps feed into later steps as inputs.

If the description is ambiguous, resolve it conservatively: create a step, mark the field "{{PLACEHOLDER}}", and add a comment in the step's name field noting the assumption.

Step 2 — Map each step to a named tool and action

For every step identified in Step 1, assign:

• tool: the named integration (use the canonical service name in snake_case, e.g., google_sheets, hubspot, slack, sendgrid, clearbit, salesforce, jira, github)
• action: the specific API operation (e.g., crm_create_contact, slack_post_message, send_email, http_request)
• inputs: the parameters the action requires, using {{trigger.row.field_name}} for trigger data, {{step_N.outputs.field}} for upstream step outputs, and {{env.VAR_NAME}} for secrets/config

Use on_error values from this set: halt_and_notify (hard failure, stop the run), continue (soft failure, proceed), continue_with_empty (treat output as empty and continue), retry_3x (retry up to 3 times before halting).

Step 3 — Construct the JSON automation definition

Emit a single JSON object with this top-level shape:

{
  "name": "<human-readable workflow name>",
  "version": "1.0.0",
  "description": "<one-sentence summary>",
  "trigger": { "type", "source", "event", "config": { ... } },
  "steps": [
    {
      "id": "step_N_<slug>",
      "name": "<human-readable name>",
      "action": "<action_type>",
      "tool": "<tool_id>",
      "depends_on": ["<step_id>", ...],
      "condition": "<optional expression — omit if unconditional>",
      "inputs": { ... },
      "outputs": { ... },
      "on_error": "<policy>"
    }
  ],
  "conditions": [
    {
      "id": "cond_<slug>",
      "description": "<plain-English explanation of the condition>",
      "expression": "<template expression>",
      "applied_to_steps": ["<step_id>", ...]
    }
  ],
  "tools": [
    {
      "id": "<tool_id>",
      "type": "<trigger_source | enrichment | crm | notification | email | storage | ...>",
      "auth_env_var": "<ENV_VAR_NAME>",
      "docs": "<canonical API docs URL>"
    }
  ],
  "env_vars_required": ["VAR_1", "VAR_2", ...],
  "execution_note": "This is a scaffold definition only. Wire it to n8n, Temporal, or Zapier to execute."
}

The output must be pure valid JSON — no Markdown fences, no prose before or after, no comments inside the JSON.

Step 4 — Self-check before emitting

Before outputting, verify:

• Every step referenced in conditions[].applied_to_steps has a matching id in steps[]
• Every depends_on entry references a real step id defined earlier in the array
• Every {{env.VAR}} used in inputs is listed in env_vars_required
• Every tool used in a step has an entry in the tools array
• The trigger.source value matches the id of a tool in tools[] (type trigger_source)
• The document is syntactically valid JSON — no trailing commas, no single-quoted strings

If any check fails, fix it before emitting.

Example

See examples/input.md for a sample natural-language workflow request (Google Sheets → Clearbit enrichment → HubSpot contact + deal → Slack notification → conditional VP email).

See examples/output.md for the complete JSON automation definition that faithfully encodes that workflow.

The example demonstrates:

• A row-added trigger on Google Sheets
• A continue_with_empty enrichment step (Clearbit may not match every domain)
• Two sequential CRM steps with depends_on chaining
• A Slack notification with inline template expressions combining trigger data and upstream step outputs
• A conditional email step gated on estimated_deal_size > 50000
• Full tools[] manifest and env_vars_required list for developer handoff