Copy That References Something Real

Outcome: an opening-line column driven by an evidence field, with banned phrases enforced, an abstain path that fires, and every line passing the interchangeability test.

Surface: App and MCP server
Level: Intermediate
Uses: AI copy column
Credits: ~1 per row
Prerequisite: Lessons 09 and 12

The test

Could this sentence have been sent to a different company?

If yes, it is decoration, not personalization — and the prospect can tell, because they have received a hundred of them. That single question decides whether a line ships.

Everything below is in service of passing it.

Rank the evidence

Build one evidence column that picks the strongest available fact per row, falling down this order:

Rank	Evidence	Why it works
1	A dated event — funding, launch, exec hire, job post	Specific and explains why you are writing now
2	Something they published — a quote from a post, a claim on a page	Proves you read their material
3	A stack fact with a consequence	Shows you understand their situation, not just their logo
4	A research verdict about their model	Understanding of the business
5	Firmographics	Everyone has this. Context only, never the hook.

The evidence column is a free formula. It does most of the work in this lesson — a good evidence field makes an average prompt produce good lines, and no prompt rescues a bad one.

The prompt


Write the opening line of a cold email to {{first_name}}, {{title}} at {{company}}.

Use only this evidence: {{evidence}}

Rules:
- One sentence, under 25 words.
- Reference the evidence specifically. Do not generalize it.
- Do not open with "I noticed", "I came across", "I saw that",
  "I was impressed by", or "As a fellow".
- Do not compliment the company.
- Do not mention our product or ask for a meeting.
- If the evidence is empty or vague, return exactly: SKIP

Return the sentence only.

“Do not mention our product” is deliberate and it is the instruction people remove first. The opening line has one job: prove the message was written for this person. The pitch that follows is the same for everyone, which means a human should write it once rather than a model writing it 500 times slightly differently.

Why the abstain path matters most

Asked to personalize with nothing to work from, a model will not fail. It will produce a fluent, confident sentence built out of nothing — a compliment about a project that does not exist, a reference to growth it inferred from a headcount number.

A prospect reading that notices immediately, and the message lands worse than a plainly generic one would have.

SKIP is what prevents it. Check that it fires: if no row returns SKIP, the abstain path is not working, and you are shipping invented personalization somewhere in the list.

The quality gate

Four free formula columns, checked before export:

Check	Rejects
8–30 words	Truncated output and paragraphs
No banned phrase	Prompt drift
Contains a distinctive token from the evidence	Lines that ignored the evidence
Not `SKIP`	Rows without usable evidence

Then read 20 rows by hand. The formulas catch shape; only reading catches tone, and tone is what gets a reply.

Do this now

Build the evidence column

Formula, ranked fallback. Count how many rows end with rank 1 or 2 evidence — that number predicts your reply rate better than the prompt does.

Write the prompt

Use the template; adjust the banned phrases to what you personally find grating.

Run on five rows and read aloud

Aloud matters. Bad AI copy reads fine on screen and sounds wrong spoken.

Apply the interchangeability test

Rewrite the prompt, not the output, for any line that fails.

Add the four gate columns

Check SKIP fires

Deliberately run it on a row with no evidence.

Widen to gated rows only

Then read 20 outputs before anything exports.

Check your work

Every line references a fact specific to that account
No banned phrases survive
SKIP appears in the output
You read 20 lines by hand, aloud
Rows with rank 1 or 2 evidence are counted

Where this breaks

Generating copy for every row regardless of evidence turns a personalized campaign into a worse generic one. The rows with nothing to say produce the most confident-sounding sentences, because the model fills the gap with plausible invention — and those are the messages that damage the relationship rather than merely failing to convert. Gate on evidence before the copy column, not after.

Further automation

Regenerate copy when the evidence changes rather than on a schedule: a new funding round or a new job post produces new evidence, a new line, and a fresh reason to re-enter the queue. That event-driven loop is what Signals & ABM and Automated Outbound build.

Next lesson

15 — Multi-step message sequences, writing follow-ups that add something instead of repeating the first touch.

Reference for this lesson: AI agents, Outreach integrations, SyncGTM 101 — writing copy, Automated Outbound.