Claude Code Outbound

A set of Claude Code skills that run cold-email outbound for a client, from onboarding to a draft campaign in your sending tool, with credit gates and quality gates so nothing gets spent or sent by accident.

The flow in one line: onboard the client, pick one ICP, search Apollo for free, approve a 10-person sample, enrich only what you approve, verify the emails, write the copy, spin it, lint real renders, and upload a draft to Smartlead or ManyReach. You press Start.

How it works

#	Stage	Skill	Cost	Your gate
1	Profile (scrapes the site first)	icp-onboarding	0	confirm the profile
2	Free offer	lead-magnet-brainstorm	0	pick one
3	Campaign ideas	campaign-strategy	0	pick campaigns
4	Search	apollo-list-builder (phase 1)	0 credits	none
5	10-sample confirmation	apollo-list-builder	0	approve, adjust or skip
6	Enrich	apollo-list-builder (phase 2)	1 Apollo credit per contact, capped	approved in step 5
7	Verify	email-verification (asks which service: MillionVerifier or Reoon, then No2Bounce)	verifier credits	pick the service
8	Copy	campaign-copywriting (4 confirmation steps)	0	approve each step
9	Spam check	spam-word-checker	0	none
10	Spintax	super-spintax (and smartlead-spintax for Smartlead syntax)	0	none
11	Upload as draft	smartlead-campaign-upload-public, smartlead-api, or manyreach-campaign-upload	0	you review and press Start

outbound-autopilot is the orchestrator that runs all of this, including several campaigns in one pass with a single review.

What's inside

skills/
  outbound-autopilot/              orchestrator + quality gates + render lint + draft-upload script
  apollo-list-builder/             free search, sample gate, capped enrichment (TypeScript scripts)
  email-verification/              MillionVerifier or Reoon (your choice) then No2Bounce, scripts/verify.ts + offline tests
  manyreach-campaign-upload/       ManyReach upload notes and variable rules
  super-spintax/                   spintax rules (20+ subject and opt-out variants)
  icp-onboarding/  lead-magnet-brainstorm/  campaign-strategy/  campaign-copywriting/
  spam-word-checker/  smartlead-spintax/  smartlead-api/  smartlead-campaign-upload-public/
examples/client-profile.example.yaml   starter profile (one ICP per profile)
tests/selftest.sh                       proves the render lint passes clean copy and catches bad data
install.sh                              links the skills into ~/.claude/skills
.env.example                            the keys the skills read

Setup

1. Requirements: Claude Code, Node 22 or newer (scripts run with --experimental-strip-types), Python 3.

2. Run the installer, then fill in your keys:

   ./install.sh
   

   This links each skill into ~/.claude/skills (it never overwrites an existing skill) and creates .env from .env.example.

3. Edit .env:

   Key	Used by
   APOLLO_API_KEY	apollo-list-builder
   MILLIONVERIFIER_API_KEY	email-verification (if you use MillionVerifier)
   REOON_KEY	email-verification (if you use Reoon)
   NO2BOUNCE_KEY	email-verification (second opinion on risky addresses)
   SMARTLEAD_API_KEY	Smartlead upload (SMARTLEADS_API_KEY is also accepted)
   MANYREACH_API_KEY	ManyReach REST path (the Manyreach MCP connector can be used instead)

   Keys are never stored in the skills. .env is git-ignored.

4. Start a new Claude Code session and run /outbound-autopilot.

Run scripts that need keys with node --env-file=.env ... so they read .env.

Email verification choices

The skill asks which services you use and stores the answer in the client profile:

Choice	Keys needed	Behavior
MillionVerifier + No2Bounce (recommended)	MILLIONVERIFIER_API_KEY, NO2BOUNCE_KEY	ok is kept, catch-all and unknown go to No2Bounce, invalid and disposable are removed
Reoon + No2Bounce	REOON_KEY, NO2BOUNCE_KEY	original routing, unchanged
MillionVerifier only	MILLIONVERIFIER_API_KEY	only ok is kept; risky addresses are written to a .held.csv for review, never sent

Offline tests for the routing: node --experimental-strip-types --test skills/email-verification/tests/verify-routing.test.ts.

Safety rules built in

• Credits: Apollo search costs nothing. Enrichment refuses to run without --max-credits and --approved. Each enrichment batch is saved so a crash never re-spends credits.
• One ICP per profile and campaign. Never combine ICPs; buyers, pains and proof differ per segment.
• Never starts a campaign. Uploads are drafts with no inboxes attached. You pick the sender and press Start.
• Identity gate: emails are written and signed as the client, with a spun signature block. Never the agency or the inbox owner.
• Render gate: render_lint.py and render_lint_sl.py render full emails with real lead values and check for leftover variables, banned words, a/an errors, doubled words, spacing, ALL CAPS, lowercase sentence starts and long subjects. Run tests/selftest.sh to confirm the lint itself works.
• Per-lead variable gate and company-name gate: per-lead lines must come from the lead's own site, and company names have banned words stripped.
• Test send: send a test to your own inbox and review it before enrolling or starting anything.

The full list is in skills/outbound-autopilot/SKILL.md, including a log of mistakes that became rules.

Quick start

/outbound-autopilot
> "New client: https://example.com. One ICP: <segment>."

Typical order: profile, one ICP, offer, five campaigns ranked by relevance to the offer, free Apollo search and 10 samples, your approval and credit cap, enrichment, verification, copy, spintax, lint, draft upload, your test send.

This is the beginning of the README. Read the rest at the original. Read the full README ↗