Agent 05 · Signal-Based Outbound
Sequence Composer: AI personalized outreach that isn’t a template with merge tags
The agent that turns a fired signal + enriched account profile into outreach that reads like a thoughtful rep wrote it — because the substance is real, not invented.
Cold outreach personalization has a ceiling, and most teams hit it: “I noticed you raised Series B — congrats!” The prospect noticed the same thing in 14 other inboxes that week. Part of the complete AI GTM engine guide.
What Sequence Composer does
Takes the structured signal payload (Intent Watcher) + the enriched account profile (Account Mapper) and generates a sequence: 4–6 touches across email + LinkedIn over 14–21 days, each touch referencing real facts about the account, with the value framing escalating across touches.
The first email is not a template. It is a generated email grounded in one specific signal and one specific account fact. The second touch acknowledges no reply and pivots framing. The third introduces a different angle. The compose pattern is the IP — not the prompt.
Why this matters
Reply rates on signal-grounded outreach run 3–10× above template outreach in the same ICP. The difference is not cleverness — it is that the prospect can immediately tell the email is about them, not about the sender’s pipeline.
The unfortunate truth: most “AI personalization” tools generate emails that pattern-match like AI. Sequence Composer avoids that because the source material (the signal) is concrete and verifiable. Concrete inputs produce concrete outputs. We run it on Anthropic Claude with per-ICP prompt scaffolding.
Stack & integration
Anthropic Claude for generation. Custom prompt scaffolding tuned per ICP. Smartlead or Lemlist for sending. n8n or Make.com for the orchestration. Output is structured so the Inbox Operator agent can route across the warmed inbox pool.
Common failure modes
- Asking the LLM to write the whole email from scratch instead of assembling from concrete facts.
- Same prompt for every ICP (ignores objection patterns, tone differences).
- Generating sequences that don’t pivot framing across touches (just “bump” emails).
- No tone guardrails (the LLM defaults to corporate-bland or aggressive-cringe).
FAQ
What LLM does Sequence Composer use?
Anthropic Claude as the default. Prompts are tuned per ICP. We don’t use generic OpenAI prompts off-the-shelf because the quality difference at high volume is meaningful and the cost difference at our volumes is not.
How is this different from Apollo's AI writer or Lavender?
Apollo and Lavender write FROM a template + a few merge fields. Sequence Composer writes FROM the specific buying signal that fired in Intent Watcher plus the account profile from Account Mapper. The personalization references concrete public facts, not 'I noticed you’re a leader in X.'
Won't prospects spot it as AI-written?
They will if the signal is weak. Strong signal-grounded outreach reads like a thoughtful rep wrote it because the substance is real. The AI is doing the assembly, not the inventing.
How many touches per sequence?
Typically 4–6 touches over 14–21 days. Each touch can deepen the angle (LinkedIn DM, follow-up email, voice note, second email with new value framing). Sequence shape is configured per ICP.
Sources & further reading
- [1]Anthropic — Claude
The default LLM Sequence Composer uses for signal-grounded generation.
- [2]Gartner — Future of Sales
Research on the shift toward digital, buyer-centric B2B engagement.
Ship Sequence Composer with the rest of the stack
14-day sprint. Code in your repo. Live demo on your data, first call.
Book live demo on your dataRelated: Lead Sourcer · Reply Triager · Inbox Operator