derivatives: outreach-sequencer

Drafts only; nothing auto-posts. Numbers trace to filled-intake-B.md.

linkedin

draft 1: closer

Cold email didn't fail at GoStudent's 100-plus tutoring centres for lack of tooling.

It failed for lack of a copywriter. Managers tried the channel, produced sequences they couldn't judge, and quietly gave up. The licences were fine. The words were the problem.

Every tool in the standard stack (Smartlead, Instantly, Lemlist, Apollo, Clay) assumes a competent operator. It hands you infrastructure. The missing layer was expertise.

So instead of another sending platform, we built the copywriter into a pipeline: AI research, one human approval gate, generation, and a deterministic last mile for the part that must never be wrong.

The full build story, including the feature that could have ended up in a newspaper:

{{POST_URL}}

draft 2: architect

The most important function in this build throws exceptions on purpose.

Merge fields must come out in each manager's exact sending-tool syntax. One wrong token breaks a campaign inside the one tool a non-technical manager can't debug.

So the model never writes tool syntax. It emits canonical markers only. A pure function substitutes the manager's real syntax at the end, and on any unknown token it throws. Zero wrong tokens becomes a contract instead of a hope.

Same philosophy on the risky part: persona research wanted a "likely income" field. It became a hedged affordability band, schema-validated at two boundaries, visible only inside the human approval gate.

Draft with a model. Approve with a human. Make the last mile deterministic.

{{POST_URL}}

draft 3: operator

The persona brief for this tool was one blunt sentence: someone who struggles to connect WhatsApp Web to a phone.

That person runs a tutoring centre. Outreach is one duty among ten. Every sending tool on the market assumes they can write audience strategy, personas, copy and token syntax in their local language. They can't, and they shouldn't have to.

The system we built makes them the editor, not the author. Their whole job: describe the centre once, pick an audience, approve the research, read the emails, export to their own Drive.

The interface even gets quieter as they get competent: coaching that's dismissible and state-aware, because help on run 1 must not patronise on run 50.

How it works, and what it deliberately refuses to do:

{{POST_URL}}

newsletter

Subject: The tools were fine. The copywriter was missing.

Body: GoStudent's offline network of 100-plus centres had every sending tool and a dormant cold-email channel. The gap was never infrastructure: centre managers, non-marketers with ten other duties, couldn't produce or judge the sequences the tools assumed they could. We built the missing expertise into a pipeline: AI research, one human approval gate, and a merge-field injector that throws on any token it doesn't recognise, so campaigns can't silently break. The build story, including its near-miss with a reputational landmine: {{POST_URL}}

sales snippet

If a channel sits dormant in your company even though the tools are paid for, the missing layer is probably expertise, not software. We build that layer in: AI drafts grounded in research, one human gate where judgement matters, and a deterministic last mile so the step that must never be wrong cannot be. Managers keep the tools they already use.

metric: template path under 100 ms, one database transaction, zero model calls (measured in build)

x thread

1. GoStudent: 100+ tutoring centres, every sending tool licensed, cold email dead. The missing piece wasn't software. It was a copywriter.
2. The user persona, verbatim from the build docs: someone who struggles to connect WhatsApp Web to a phone. Now ask them for a 5-step sequence in their local language with correct merge tokens.
3. Smartlead, Instantly, Lemlist, Apollo, Clay: all assume a competent operator. Infrastructure was never the gap. Expertise was.
4. So the build inverted: no sending, managers keep their tools. AI persona research → one mandatory human gate → generation → deterministic merge-field injection → export to their own Drive.
5. The dangerous part: research wanted "likely monthly income". It shipped as a hedged affordability band, schema-validated at two boundaries, shown only inside the approval gate. Structure beats politeness.
6. The trustable part: the model never writes tool syntax. It emits canonical markers; a pure function injects the real syntax and THROWS on unknown tokens. A contract, not a hope.
7. Honest limits: business results not yet measured; sign-off gate is 10 samples across 7 languages; concurrency capped at 5 by the vendor plan. Design truths today, numbers next. {{POST_URL}}