Playbook · Outbound

Playbook: Run cold outbound with agents without burning your domain

Mine real calls for language, enrich with signals, write one email per row, and keep deliverability guardrails and a human hand-off in place.

Teardowns: 3Stack: Claude Code · Clay · Instantly / ReplyConfidence: Low to medium
Alejandro Arce · Ilya Azovtsev · Shannon Torcato Verified Oct 16, 2026 Markdown copy

Who it’s for: B2B founders and small agencies who want agents to do the research, writing and reply triage in cold outreach.

Tools these builders use: Claude Code, YAML rules, Postmark or SendGrid, a warmup tool, HubSpot or Pipedrive, Calendly (Alejandro Arce); Claude, Clay, Jason AI (Reply) (Ilya Azovtsev); Claude Code, Clay, Apollo, Instantly, n8n, Slack, HubSpot, Cal, HeyReach and several enrichment and signal tools (Shannon Torcato).

The steps

  1. Prove the offer by hand first. Alejandro’s rule: get to a 10%+ reply rate manually before automating. Source: Alejandro (teardown, original)

  2. Pull the language from real sales calls. Ilya had Claude analyze 3 discovery calls and extract pain points like “net new is hard” and “inbound hit a ceiling”, then write the whole sequence in the prospects’ own words. Source: Ilya (teardown, original)

  3. Define the ICP and build an enriched list. Alejandro’s Claude Code setup pulls prospects, adds recent news, hiring posts and tech-adoption signals, and validates emails with a catch-all checker, producing a 200-500 row CSV. Ilya scraped the HubSpot Partner Directory and enriched it in Clay. Shannon’s list agent tiers accounts in Clay and fills missing contacts through a waterfall of providers. Sources: Alejandro (teardown, original); Ilya (teardown, original); Shannon (teardown, original)

  4. Personalize from data, with examples in context. Alejandro fills a templated structure from the enrichment data with 50+ good and bad examples in context, and anything that reads as forced gets flagged and rewritten. Source: Alejandro (teardown, original)

  5. Write the follow-up rules down. Alejandro’s YAML: no open in 3 days gets a new subject line; opened but no reply in 5 days gets a value-add email; “not interested” is removed; interested gets a booking link. Source: Alejandro (teardown, original)

  6. Set deliverability guardrails before sending. Alejandro uses a dedicated domain, warms up from 10 emails a day over 2-3 weeks, sets SPF, DKIM and DMARC, and pauses automatically if bounces pass 3%. Shannon’s deliverability agent warms inboxes, rotates sending accounts and pauses flagged inboxes. Sources: Alejandro (teardown, original); Shannon (teardown, original)

  7. Let the agent run the campaign tool and watch replies. Ilya’s Claude added contacts, launched the sequence in Jason AI (Reply), monitored deliverability and flagged hot leads: “I just showed up to book the calls.” Shannon’s reply agent classifies each reply by intent and drafts responses with the thread context. Sources: Ilya (teardown, original); Shannon (teardown, original)

  8. Hand engaged prospects to a person and the CRM. Alejandro lets a human take over once a prospect engages. Replies create a CRM lead with the campaign source, and anyone who doesn’t book within 48 hours gets a nudge. Source: Alejandro (teardown, original)

  9. Log wins and losses back into targeting. Shannon’s learning agent logs booked meetings to the CRM, refines targeting by which segments converted and generates new angles from win and loss patterns. Source: Shannon (teardown, original)

Postby’s suggestion: keep a human reading replies before anything is sent back. Shannon’s post says meetings get booked “without me being involved at all” and doesn’t say whether drafted replies are reviewed; Alejandro hands engaged prospects to a person. Postby would follow Alejandro’s version.

Human approval points

  • Alejandro: a human takes over once a prospect engages.
  • Ilya: he booked the calls himself; the post doesn’t describe reviewing copy before launch.
  • Shannon: none described.

Common failure points (as the builders report them)

  • Alejandro names the two things that kill campaigns: poor personalization and inconsistent follow-up.
  • Bounce rate: his automatic pause at 3%.

Caveats carried over

  • Alejandro: client results (450 prospects, 62 replies, 18 calls, 4 deals) are self-reported, Unverified · independently. Scraping LinkedIn may conflict with LinkedIn’s terms; this is his method, not an endorsement.
  • Ilya: one LinkedIn post, 60 contacts, one week, no raw data. 17 calls marked “(so far)”. The setup is offered via comment-for-DM. X: Unverified.
  • Shannon: an architecture outline. No results, prompts or configs. Low confidence. X: Unverified.
  • Not included as a step source: Tim Yakubson, because his workflow is gated behind a comment and we haven’t seen it.

Examples

Get one teardown a week.

Same format, a new operator every issue. Double opt-in.