Scott Krukowski Scott Krukowski

Case Study · AI Workflow

Operating

From ICP to Gmail drafts in one command.

A five-stage pipeline — ICP spec → Exa web research → contact enrichment → Claude drafting → Gmail drafts — that replaces the research-to-outreach grind. Human approves and sends. Drafts only, never auto-send. Zero retainer.

Workflow Pipeline

ICP → Research → Enrichment → Draft → Gmail

Step 01 · Scope

Target ICP

  • Role · geo · size
  • Intent signals

Step 02 · Discover

Exa Research

  • Web search
  • Context signals

Step 03 · Enrich

Apollo / Clearbit

  • Verified email
  • Title + company

Step 04 · Draft

Claude Drafting

  • Personalize per prospect
  • Research-backed hook

Step 05 · Stage

Gmail Drafts

  • Human reviews
  • Human sends
Copper: data / research Teal: LLM + staging Human-approved send · drafts only

Demo Reel · Coming

Pipeline Run — End to End

30-second recording: kick off the skill in Claude Code, watch Exa surface prospects, enrichment fill contact fields, Claude draft, Gmail drafts land.

Placeholder

Sample Draft · Redacted

To: [buyer@prospect-co.example]

Subject: Question on [vertical]-grade adhesive sourcing

[First name] —

Saw your team's recent [project / product launch / announcement discovered by Exa research]. Congrats on the [specific detail] — the [technical angle] stood out.

We run a [specialty manufacturing] function that's been quietly doing [specific capability tied to their need] for similar [vertical] players. One thing we help unlock is [differentiated benefit].

Worth a short call next week? I'll bring a short prepared comparison against your likely current spec.

— [Sender]

Example shape · one draft per prospect, personalized per research pull

The problem

Cold outbound done well is the grind most teams don't do well. It requires fresh research (what are they actually building?), enriched contact data, a real reason to reach out, and personalization that doesn't read like a template. Doing it right for ten prospects is a full day. The obvious fix — hire an SDR or an agency — burns $3–5K/month and still produces templated output. The wrong fix — scripted spray-and-pray — burns deliverability and brand.

The pipeline

Five stages, run as a single Claude Code skill. Input: an ICP spec (role, geo, size, verticals, intent signals). Output: ten to fifty Gmail drafts, each personalized against fresh research.

  1. Scope. Skill reads the ICP spec file from the workspace.
  2. Discover. Exa web search pulls candidate companies + context signals (recent news, product launches, growth indicators).
  3. Enrich. Apollo / Clearbit fills in title, verified email, company size.
  4. Draft. Claude writes a personalized email using the ICP spec, the research signal, and the enriched contact — not a template, a real paragraph.
  5. Stage. Gmail API creates the draft in my drafts folder. Nothing sends without my review and click.

Design decisions

  • Drafts, not sends. Every email enters Gmail as a draft. Reviewing is a 20-second scan per message. No autonomy without accountability.
  • Research before draft. Personalization comes from Exa finding a real, recent signal — not from a first-name mail-merge pretending to be research.
  • Skill, not script. Packaged as a Claude Code skill so it's version-controlled, auditable, and composable with other workflows in the same workspace.
  • Portable ICP. Swap the ICP spec file and the entire pipeline retargets. No per-vertical rebuild.

Outcomes

  • Compresses a full day of manual prospect research + drafting into roughly 10 minutes
  • Replaces ~$3–5K/month in outsourced SDR / VA outreach tooling
  • Human-in-the-loop by design — every email enters Gmail as a draft, never auto-sends
  • Portable — same skill runs against any ICP definition, not a single-vertical hardcode

Stack

Claude Code SkillsExa APIApollo / Clearbit (enrichment)Claude APIGmail APIPython
← Back to Portfolio See Resume