/auto-research-public

name: auto-research-public
description: Autonomous cold email campaign launcher. Takes one target company domain, scrapes their website, generates an ICP with Claude, pulls matching leads via Prospeo, enriches emails + company descriptions, personalizes each lead with parallel Claude Code Task sub-agents (A/B/C variants), and uploads as a live Smartlead campaign. Uses local JSON files for state (no database required). Use for automated daily campaign launches after you have an initial `client-profile.yaml` from /icp-onboarding. Triggers on "auto-research", "launch an automated campaign", "daily campaign", "run the research loop".

Auto Research (Public)

Automated end-to-end campaign launcher. Feed it one target company domain, get back a live Smartlead campaign with per-lead personalization — in about 20 minutes.

This is the beginner-friendly version of the GEX internal `auto-research-v2`. All state lives in local JSON files; no Supabase, no Trigger.dev.

What you get

• **Input:** one target company domain + your `client-profile.yaml`
• **Output:** a running Smartlead campaign with:
• 200-1,000 leads (depending on targeting tightness)
• Per-lead personalization: 9 custom variables (situation, value, CTA × 3 variants)
• A/B/C subject + body variants tested in parallel
• Campaign assigned to your available inboxes
• Schedule: Mon-Fri 8am-5pm your timezone

Prerequisites

Before running:

• [ ] `client-profile.yaml` exists (run `/icp-onboarding` if not)
• [ ] `SMARTLEAD_API_KEY` in env
• [ ] `PROSPEO_API_KEY` in env
• [ ] `MILLIONVERIFIER_API_KEY` in env (for email validation)
• [ ] At least 20 Smartlead inboxes tagged "active" (run `/smartlead-inbox-manager` first)
• [ ] At least 1 campaign template in Smartlead (or the script creates a fresh one)

The orchestration (Claude Code runs this)

Unlike the other skills, this skill orchestrates through the Claude Code conversation itself — Claude does the reasoning (ICP generation, copy writing, personalization), and phase scripts do the heavy API I/O. This is the pattern from the GEX v2 internal.

Phase 1: Scrape the target company

npx tsx scripts/phase-scrape.ts --domain=<target.com> --out=/tmp/auto/scrape.json

Output: JSON with `domain` + text content from homepage, /about, /product, /pricing, /customers.

Claude reads the output and writes a short analysis to `/tmp/auto/company-analysis.md`:

• What the company does
• Who their likely customers are
• Social proof signals
• Potential angles for outreach

Phase 2: Claude generates ICP filters

Reading `/tmp/auto/scrape.json` + `~/cold-email-ai-skills/profiles/<slug>/client-profile.yaml`, Claude writes Prospeo filters to `/tmp/auto/filters.json`:

{
  "job_titles": ["VP Marketing", "Head of Marketing", ...],
  "seniorities": ["Vice President", "Head", "Director"],
  "industries": ["Software Development", "Financial Services"],
  "company_size_min": 50,
  "company_size_max": 500,
  "countries": ["US"],
  "excluded_industries": ["Religious Institutions", "Government Administration"]
}

Claude MUST use exact Prospeo industry names from `~/cold-email-ai-skills/skills/icp-onboarding/references/prospeo-industries.md`.

Phase 3: Prospeo search

npx tsx scripts/phase-prospeo.ts --filters-file=/tmp/auto/filters.json --max-leads=1000 --out=/tmp/auto/leads.json

Output: JSON with `leads` array. Each lead has: `first_name`, `last_name`, `email` (may be empty), `linkedin_url`, `job_title`, `company_name`, `company_domain`, `company_industry`, `company_headcount`, `company_description`.

Phase 4: Email waterfall + description enrichment

npx tsx scripts/phase-enrich.ts --leads-file=/tmp/auto/leads.json --out=/tmp/auto/enriched.json

The script: 1. Checks each lead for email; if missing, hits Prospeo's `enrich-person` endpoint 2. If company_description is thin (<50 chars), scrapes company_domain homepage 3. Runs MillionVerifier on every candidate email 4. Writes enriched leads (only those with valid email) to output

Expect hit rates:

• Retail/SMB: ~99% email found
• B2B tech: ~65-80%
• Healthcare/public sector: ~25-40%
• MV rejection: ~20-30% of found emails

Phase 5: Copy writing (Claude)

Claude generates 3 copy variants (A, B, C) and writes to `/tmp/auto/variants.json`:

[
  {
    "variant": "A",
    "subject": "<concrete, specific, <60 chars>",
    "angle": "<pain observation>",
    "body_template": "Hi {{first_name}},\n\n{{situation_line_a}}\n\n{{value_line_a}}\n\n{{cta_line_a}}\n\n%signature%\n\nP.S. If this isn't relevant, just let me know and I won't reach out again."
  },
  {
    "variant": "B",
    "subject": "<different angle>",
    "angle": "<compliment + transition>",
    "body_template": "..."
  },
  {
    "variant": "C",
    "subject": "<third angle>",
    "angle": "<question>",
    "body_template": "..."
  }
]

Rules Claude MUST follow (run `/spam-word-checker` on output):

• No em dashes (—). Use commas or periods.
• No "leverage", "synergy", "solutions", "world-class", "cutting-edge".
• Body: 50-90 words max.
• Subject: under 60 chars, specific, no clickbait.
• End with `%signature%` on its own line.

Phase 6: Personalization via Task sub-agents

Claude fans out to parallel Task sub-agents (one per variant × batch of 20-30 leads). See `/personalization-subagent-pattern` for the full pattern.

Per lead, each sub-agent writes to `/tmp/auto/personalization-<batch>-variant-<X>.json`:

[
  {
    "lead_id": "...",
    "situation_line": "One sentence about the company.",
    "value_line": "One sentence connecting to our offer.",
    "cta_soft": "One soft ask sentence."
  }
]

After all sub-agents finish, Claude merges by lead_id into `/tmp/auto/personalized.json`. Each lead now has 9 personalization fields.

Phase 7: Upload to Smartlead

npx tsx scripts/phase-upload.ts \
  --leads-file=/tmp/auto/personalized.json \
  --variants-file=/tmp/auto/variants.json \