MCP Master

Leadspicker Master Pipeline (MCP)

End-to-end orchestrator that takes a project from "leads loaded" to "outreach sequence ready". Runs five phases:

1. Setup — resolve project, gather business context.
2. Enrichment — populate the data needed for everything downstream.
3. Classification — score / filter by ICP fit and any boolean criteria.
4. Personalization — generate hooks, icebreakers, salutations.
5. Outreach sequence — build the email / LinkedIn / multi-channel graph.

This skill does not duplicate the detailed how-to of each phase. For prompt templates, column tables, BASHO rules, vocative grammar, and so on, request the dedicated skill by name when you start that phase:

• mcp-data-enrichment — built-in + AI columns, dependency map, lean defaults.
• mcp-classification — boolean / string / scored prompts, output mode decision.
• mcp-personalizer — 14–16-word hooks, icebreakers, multi-language salutations & vocatives.
• mcp-outreach — sequence step types, BASHO framework, multi-channel graph.

Mention them by name in your replies (e.g. "loading mcp-classification for this phase") — the user / runtime decides whether to load the full skill content.

MCP tools used directly by this skill

For navigation only — each phase mostly delegates to the skill above, which calls its own tools.

⚠️ Credit discipline (applies across all phases)

Every enrichment / AI column costs credits. Default to a lean minimum. Only add optional columns when explicitly triggered by the user's persona or criteria.

• NEVER default to li_company_employee_count (exact count). Use li_company_size (range).
• NEVER default to li_latest_posts. It is the most expensive enrichment.
• For each phase, present the planned columns, let the user approve, and explicitly call out what's NOT being enriched and why ("saves credits").

For full credit-discipline rules and the lean default column set, request mcp-data-enrichment.

Phase 0 — Setup

Resolve the target project and collect business context.

1. Display format: always render projects as {name} (#{id}).
2. Resolve the project:
   • User named one → mcp__leadspicker__list_projects(search_query="{name}", limit=5). If multiple match, ask which.
   • Otherwise → mcp__leadspicker__list_projects(order_by="last_active", limit=20) and ask the user to pick.
3. Gather, in one consolidated question if possible:
   • What you sell / offer. Product, value prop, problem solved.
   • ICP. Industry / vertical, company size, role / seniority of the buyer, geography.
   • Language. Default English; ask if non-English (sequences + personalization need it).
   • LinkedIn Premium? Determines connection-request text and InMail availability.
   • Sender name for LinkedIn messages (LinkedIn has no auto-signature).
4. Inspect the project:
   • mcp__leadspicker__list_contacts(project_id=X, page=1, page_size=1) — total contacts.
   • mcp__leadspicker__list_magic_columns(project_id=X) — what already exists (avoid re-creating columns).
5. Show a one-line summary of what you understood, and confirm before moving to Phase 1.

Phase 1 — Enrichment

Goal: populate the data classification + personalization need, no more.

1. Request mcp-data-enrichment for column tables, dependency map, and the Lean Outreach Prep pipeline.
2. Build the lean default column set from the user's input:
   • Always: company_name, company_website, company_linkedin, person_linkedin, li_full_name, person_position, li_company_description, website_text_summary, enrich_emails, validate_emails.
   • Add only when triggered by the user's persona / criteria:
     • geography → person_country and/or li_company_country
     • size band → li_company_size (range; never the exact count)
     • department → matching li_company_people_*
     • company age → company_founded_year
     • personal bio / experience → li_about_me, present_experience
     • icebreakers from posts → li_latest_posts (confirm first — expensive)
3. Show the enrichment plan in three sections: Always, Added because you mentioned X, NOT enriching (saves credits).
4. After approval, launch via mcp__leadspicker__enrich_contacts per column (≈1 s delay between calls).

Checkpoint A

Enrichment is asynchronous — columns populate over minutes. Tell the user:

"Enrichment is running. Check the project in Leadspicker; once the columns are populated, say continue and I'll move on to classification."

Do not start Phase 2 until the user says continue.

Phase 2 — Classification

Goal: score and filter contacts by ICP fit. Two waves:

2a — scored relevance

For prompt structure, output-mode decision (boolean / string / scored), and variable selection, request mcp-classification.

Suggested scored columns (is_boolean=False, "Yes | score | reasoning" format):

• Position relevance — does the person's role match the ICP?
  • Variables: {{position}}
  • Prompt customized to the ICP roles the user described.
• Company relevance — does the company match the ICP?
  • Variables: {{company_name}}, {{company_website}}, {{linkedin_company_description}}, {{website_text_summary}}
  • Prompt customized to the ICP industry / vertical / size.

Run a preview (mcp__leadspicker__preview_ai_prompt) on the first 5 rows for each column. Show the user the combined preview.

Checkpoint B

"Scored classifications are running. Please check the previews / first results in Leadspicker. When you're happy, say continue and I'll add the boolean filter columns."

2b — boolean filter columns

After the scored columns finish, add boolean filter columns that wrap them:

• Position Relevant? (is_boolean=True) reading the position-relevance score (e.g. score ≥ 7 → true)
• Company Relevant? (is_boolean=True) reading the company-relevance score similarly

Plus any extra boolean filters the user asked for (e.g. "is decision maker", "targets enterprise").

These boolean columns become the condition gates of the outreach sequence in Phase 5.

Checkpoint C

"Boolean filters are processing. Once they finish in Leadspicker, say continue and I'll move on to personalization."

Phase 3 — Personalization

Goal: short, natural, language-correct opener fragments — never full messages.

For prompt rules, language tables, vocative grammar, and auto-generation rules, request mcp-personalizer.

Default for the user's language

Icebreaker is only generated if li_latest_posts was enriched in Phase 1.

Checkpoint D

Preview at least the website hook on first 5 rows (mcp__leadspicker__preview_ai_prompt). Show the combined table to the user. After approval, launch all personalization columns (mcp__leadspicker__enrich_contacts, ≈1 s delay between calls).

"Personalization columns are processing. Once they finish in Leadspicker, say continue and I'll build the outreach sequence."

Phase 4 — Outreach sequence

Goal: turn approved data + personalization into the sequence graph.

For step types, BASHO framework, word limits, spintax, and the multi-channel graph, request mcp-outreach.

1. Ask the user which sequence type to build: email, LinkedIn, or multi-channel.
2. Reuse the answers from Phase 0 (offer / ICP / language / Premium / sender name) — don't ask again.
3. Find boolean column header IDs:

   mcp__leadspicker__get_project(project_id=X, info_type="detail")
   

   Look in headers_data for entries with is_boolean: true and pick their id.
4. If any boolean columns exist, ALWAYS add them as magic_column_condition steps at the root of the sequence (chained via relation_type: "yes"). Filters rolled out in Phase 2b act as the gate.
5. Use available variables when generating message text:
   • {{first_name}}, {{last_name}}, {{full_name}}, {{position}}, {{company_name}}
   • Personalization columns from Phase 3: {{website_hook}}, {{company_hook}}, {{icebreaker}}, {{salutation}}, vocative columns.
6. Apply BASHO + word limits + spintax (mcp-outreach has the details).
7. Show all messages to the user for approval.
8. Delete any existing sequence: mcp__leadspicker__delete_sequence(project_id=X).
9. Build the sequence in one call: mcp__leadspicker__create_sequence(project_id=X, sequence_messages=...). Never chain multiple create_sequence_step calls.
10. Verify: mcp__leadspicker__list_sequence(project_id=X).

Skip rules

The pipeline shrinks when the user does not need every phase:

• Project already enriched (most lean columns exist) → skim Phase 1, only add what's missing.
• No boolean classifications needed → skip Phase 2b; Phase 5 has no condition gates.
• English-only outreach without posts → Phase 3 skips icebreaker + vocative columns; only website / company hooks.
• User already has a sequence and only wants data prep → stop after Phase 3.
• User already has data + classification + personalization and only wants the sequence → jump straight to Phase 4 (mcp-outreach).

Pipeline summary template

After Phase 4 finishes, give the user a one-screen recap:

Pipeline complete ✓ — {project_name} (#{id})

Enrichment:        {N} columns ({list})
Classification:    {scored cols}, {boolean filter cols}
Personalization:   {hook / salutation / vocative cols}
Sequence:          {type} ({step count} steps, {condition gates} gates)

Next step in Leadspicker:
  • Connect the email / LinkedIn account if not yet attached
  • Set sending hours via mcp-outreach (`get_sequence_settings`)
  • Launch the sequence from the Leadspicker UI

Natural language matching

Safety

• Always show the plan before each phase and wait for explicit approval.
• Apply credit discipline — never auto-enrich "just in case".
• Use checkpoints (A, B, C, D) — wait for the user to say continue between phases that depend on async results.
• Treat each phase's detailed rules as the source of truth — request the corresponding mcp-… skill before generating prompts or sequence graphs.