Multi-Step Message Sequences
Outcome: a four-step sequence where each step has a distinct job and only the parts that must differ per row are generated.
- Surface
- App
- Level
- Intermediate
- Uses
- AI copy columns · outreach integrations
- Credits
- ~2 per row for a full sequence
- Prerequisite
- Lesson 14
Most follow-ups say nothing
“Just bumping this to the top of your inbox.” “Did you get a chance to see my last email?” “Following up on the below.”
These add no information, and a prospect who ignored message one has no new reason to read message two. Worse, they signal an automated cadence more clearly than the first message did.
The fix is structural: give every step a different job.
Four steps, four jobs
Step 1 — the specific reason
The evidence-driven opening from lesson 14, plus a short, plain statement of what you do and a low-friction ask.
Step 2 — a different angle on the same problem
Not a reminder. A second piece of evidence, a consequence they may not have considered, or a peer example. If you have nothing new, this step should not exist.
Step 3 — proof
A customer like them, with a number. Shortest message in the sequence. This is the step that most often converts, because it answers “does this work for someone in my position?”
Step 4 — the close-out
Explicitly the last message. No guilt, no “I’ll assume you’re not interested”. One line offering to close the loop, and a door left open.
Four is a workable default. Beyond about five, additional steps mostly generate unsubscribes.
Generate the minimum
Per-row generation is the expensive part, so generate only what must differ.
| Element | Generated per row? |
|---|---|
| Step 1 opening line | Yes — the evidence sentence |
| Step 1 body and ask | No — written once by a human |
| Step 2 angle | Yes — if you have second-rank evidence |
| Step 3 proof | No — selected from a small set by segment |
| Step 4 close-out | No — one version, for everyone |
That is two generated sentences per prospect, not four full emails. It costs a fifth as much and reads better, because the human-written parts are actually good.
Choosing a case study by segment is a lookup, not a generation. Classify the row’s segment in lesson 11, map segment to case study in a formula, and insert it. Free, and more accurate than asking a model to pick.
Making step 2 different
The rule: step 2 must contain a fact that was not in step 1.
Practical sources, in order of strength:
- The second-ranked evidence from your ranking in lesson 14.
- A consequence drawn from the first evidence — “you’re hiring three data engineers, which usually means someone is about to own this manually for six months.”
- A peer at a comparable company doing the thing you are describing.
If none of those exists, drop step 2 entirely and go straight to proof. A three-step sequence with three real messages beats a four-step one with a filler.
Spacing
| Step | Days after previous |
|---|---|
| 1 | — |
| 2 | 3–4 |
| 3 | 4–5 |
| 4 | 7 |
Two principles behind these numbers: nothing within 48 hours, because it reads as pressure; and total sequence length under three weeks, because past that the reason you wrote in the first place has expired — which is the whole point of a signal-driven campaign.
Do this now
Assign each step a job
Write the job in one line before writing any copy.
Write the human parts once
Body, ask, proof set, close-out. These are the same for everyone and deserve real effort.
Generate only the two sentences
Step 1 opening, step 2 angle.
Check step 2 against step 1
Row by row on ten rows. Any pair where step 2 adds nothing means step 2 should be dropped for that segment.
Map proof by segment
Formula lookup, not generation.
Set spacing
Nothing under 48 hours; sequence under three weeks.
Read one full sequence end to end
As a prospect would. Four messages, in order. This catches repetition nothing else does.
Check your work
- Each step has a stated, different job
- Only two sentences per prospect are generated
- Step 2 contains a fact absent from step 1
- Case studies are selected by lookup, not generated
- You read a complete sequence as a prospect would
Where this breaks
Generating all four messages per row produces four variations on the same sentence, because the model has one evidence field and four chances to use it. The prospect receives what reads as the same email rewritten — which is worse than an honest follow-up, and it costs four times as much. Generate the sentences that must differ; write the rest once, properly.
Further automation
Kill the sequence when the reason expires. A prospect sequenced because they posted a role that has since been filled is being contacted about something that is no longer true — and a signal-driven campaign that ignores signal decay is just a cold campaign with extra steps. Wire the window check into the sequence, not just into the list build.
Next lesson
16 — Quality control on AI output, the checks that stop bad rows leaving the table.
Reference for this lesson: Outreach integrations, Instantly, AI agents, Automated Outbound.