/plan

Design a GTM Workflow

Goal

<workflow_goal>$ARGUMENTS</workflow_goal>

If the goal above is empty, ask: "What workflow do you want to build? Describe the data flow you want to achieve."

Do not proceed until you have a clear goal.

Phase 1: Survey the Environment

Before designing anything, gather context by calling these MCP tools:

1. list_tables — See what tables already exist. The user may have existing data to build on.
2. get_connected_platforms — See which integrations are connected (HubSpot, Slack, LinkedIn, etc.). This determines which actions are available.
3. list_actions — Get the full action catalog. Note which have hasDetailedGuide: true.

If relevant tables already exist, also call get_table_schema on each to understand current fields, data types, and what's already been built.

If the user mentions specific actions or integrations, call get_action_schema to read the aiDescription for those actions now — you'll need this context for the design.

Phase 2: Design the Architecture

Based on the goal and available tools, design the workflow:

1. Identify entity types

What data entities are involved? Companies, contacts, deals? Each gets its own table.

2. Map the data flow

For each table, define the field chain in order:

• Source: Where does data come from? (LinkedIn import, HubSpot list, webhook, manual)
• Enrichment: What data needs to be added? (enrich_company, enrich_contact, email/phone enrichment)
• People finding: If the workflow needs to find contacts at companies, choose the right method based on target audience:
  • LinkedIn-heavy (tech, enterprise, B2B): li_find_people_at_company only
  • Non-LinkedIn (small businesses, non-tech, low-LinkedIn regions): custom_ai_agent with web search + JSON Schema only
  • Mixed/uncertain: both, with AI web search gated on LinkedIn isNotFound
  • See workflow-patterns.md "People-Finding Strategy" for details. Do not build both unless the audience warrants it.
• Qualification: What filtering/scoring is needed? (custom_ai_agent, formula, perplexity)
• Routing: Does data flow to other tables? (send_to_table with mode and field mappings)
• Sync: Does data go to a CRM or outreach tool? (hubspot_create_object, outreach actions)

3. Define autoRunConditions

Which fields gate on which upstream results? Apply the "filter cheap before expensive" principle:

• Free: formulas, lookups
• Cheap: enrichment (1-2 credits)
• Expensive: AI + web search (5-50 credits), findPeople (2 credits/contact)

4. Check CRM integrity

Any CRM sync MUST follow the lookup-before-create pattern. Verify that:

• Lookup fields exist before create fields
• Create fields are gated on lookup returning isNotFound
• Parent record IDs are passed through (e.g., company HubSpot ID to contacts)
• Company association rule: If the workflow pushes contacts to HubSpot after a job change or company enrichment, the plan MUST include company lookup-before-create and contact-company association. Never update a contact's company as a flat text field without also creating/linking the Company object. This means: resolve company domain → lookup company in HubSpot → create if not found → update contact with associateWithObject: true pointing to the company's HubSpot ID.

Phase 3: Present the Plan

Output the workflow architecture in this format:

Tables

For each table:

• Name and entity type
• Source (how data gets in)
• Field chain (ordered list with action key, autoRunCondition, purpose)
• Send to Table connections (if routing to other tables, specify mode and key mappings)

Data Flow Diagram

Show the table-to-table flow:

[Companies] --send_to_table(send_for_each_item)--> [Contacts] --hubspot_create_object--> HubSpot

Cost Estimate

Approximate credit cost per row flowing through the full workflow. Note which actions are free vs. paid.

Risks and Considerations

Flag any concerns: missing integrations, data quality requirements, rate limits, non-deterministic steps.

Standard risk items to check:

• Company domain availability: If the workflow enriches contacts and the enrichment may return null for companyWebsite, the plan must include an AI domain resolution step (custom_ai_agent with web search, ~4 credits per row) before HubSpot company lookup. Flag this cost in the estimate.

Testing Strategy

Define how the workflow will be validated before running on the full dataset:

1. Rung 1 checkpoints — for each table, what does a successful 1-row test look like? (Expected output values, expected row count in destination tables, expected CRM state)
2. Rung 2 scope — 10 rows through the full chain. What to verify at scale (error rate, data quality, CRM dedup).
3. Test cost — estimated credits for Rung 1 (1 row × full chain) + Rung 2 (10 rows × full chain).
4. Full-scale cost — estimated credits for all rows. This number will be reported to the user before Rung 3.
5. Rung 3 batch strategy — for tables with >100 rows, plan to use list_row_ids (with hasNotRun filter) to paginate row IDs, then batch through run_fields with custom_range (100 rows per batch).

Phase 4: Confirm and Handoff

Do NOT create any tables or fields. This command is plan-only.

Present three options to the user:

1. Build step by step — /baseloop-gtm:build — Create tables and fields one at a time, verifying each step before proceeding.
2. Build autonomously — /baseloop-gtm:lfg — Plan, build, and test (Rung 1 + Rung 2) autonomously. Pauses for your approval before running on the full dataset (Rung 3).
3. Adjust the plan — Modify the architecture before building.

Ask: "Ready to build? Choose build for step-by-step, lfg for autonomous end-to-end, or tell me what to adjust."