Lead enrichment pipeline

# Lead enrichment pipeline

You have the "lead-scorer" MCP server connected (Lead Scorer CRM — endpoint https://mcp.lead-scorer.com/mcp, authenticated with Lead Scorer OAuth). Use its tools for every read and write. Discover resource IDs with the available list/search tools; never guess or probe sequential IDs, and ask me when no discovery tool exists. Never invent data: if a tool result is empty, say so. An API key is only a manual fallback for clients without OAuth support.

## Goal
Take list <LIST_ID> from raw to campaign-ready: enriched profiles, found emails, and an AI summary per lead.

## What this costs — read before step 2
Enrichment is the only part of this chain that spends credits. Reads, scores and drafts are free.

| Call | Cost |
| --- | --- |
| `enrich_leads` | ~1 credit per lead |
| `enrich_company` | 1 credit (0 when served from cache) |
| `find_lead_contact_info` | 3 credits per email FOUND, 15 per phone found (nothing found, nothing charged) |

**Never enrich a whole list on arrival.** Pre-qualify for free first with the "ICP scoring rubric" skill (`submit_lead_score` costs nothing), then enrich only what scored as plausible ICP. Enriching blind costs around 8 credits per usable contact; pre-qualifying first brings it close to 1.

Both paid calls accept `dry_run: true`, which returns the estimate and your balance without spending anything. Above the account threshold they refuse to run without `confirm: true`, and the refusal states what it would have cost. That refusal is the feature, not an obstacle to route around.

## Steps
1. **Inventory.** `get_leads_from_list` — count who has a LinkedIn URL, an email, an enriched profile.
2. **Estimate, then enrich the pre-qualified only.** Call `enrich_leads` once with `dry_run: true` and the complete shortlist, then call it once for real with the same `lead_ids`. The backend persists one resumable run and batches provider work safely; one tool call per lead creates a burst of competing runs and must never be used. Enrichment REQUIRES the lead to have a linkedin_url — skip and report those without one. Use `enrich_company` for companies with a LinkedIn but no data.
3. **Contact finding.** Hand this to the "Contact discovery" skill: it only runs on leads confirmed as ICP that you intend to contact. An email found for someone nobody will write to is 3 credits burned — a miss is free, a useless hit is not.
4. **Company research.** Select the owned product/ICP, call `get_companies_pending_research` (optionally scoped to the company list), then read each official product site. Save a factual, source-backed site summary plus a separate analysis against that product and ICP with `submit_company_research`. Do not draft outreach in this step.
5. **Lead AI layer.** `get_leads_pending_ai_enrichment`, then for each: write a 3-4 line summary and 2-3 actionable insights (angle to open with, risk, timing) from the enriched data, and store them with `submit_lead_ai_enrichment`.
6. **Reconcile.** Enrichment data is the source of truth: if it contradicts what web research said (role changed, company pivoted), update your notes and say so.
7. **Report.** Enriched / company-researched / emails found / AI-summarized / skipped (and why).

## Hard rules
- One shortlist per `enrich_leads` call, up to 1,000 leads. Never loop or parallelize one call per lead.
- Keep the returned run ID and poll it with `get_lead_enrichment_run`; use `list_lead_enrichment_runs` to recover progress after a restarted session.
- Nothing paid runs on a lead that has not been scored first.
- Company research must cite the official site. Keep factual site summary separate from the interpretation tied to the selected product/ICP.
- Write CRM summaries, research analyses and insights in English, even with non-English sources. Preserve names and source quotations. Outreach language is selected independently.
- No outreach here. This skill ends when the list is ready.