Playbook AI Specificity

Playbook: AI Specificity Line

All rules here are best practice, not law. Override any of them when the campaign calls for it; note the best practice once and move on.

Use when: the offer is good but the email reads like it was sent to 5,000 people, and you need one sentence that names what the client would actually do for this company.

Do not use when: you want a first line about a person or an event. Those playbooks personalize the opener. This one personalizes the offer — which is a different and usually more durable move.

One-line output: specificity_line = "know the landed cost of every pool float before you price it"

1. Trigger and scope

This produces the tail of one fixed sentence:

I think we can help you {{specificity_line}}.

The whole job is one move: take the client's generic capability and say it back in the prospect's own nouns.

• A bookkeeping firm does not write "we help with margin reporting". It writes "know what each pool noodle costs to make before you price the next run".
• A list-building service does not write "we build B2B lists". It writes "build a list of every pool contractor in Texas with the owner's cell number".

Same offer, different words, and the prospect reads it as written for them.

Two things this is not. It is not a first line — it never mentions the person, their job, or anything they did. And it is not a claim about the company's problems — it never says they are struggling or missing something. It says what the work would be.

⚠️ The frame lives in the campaign copy, not inside the variable. Real campaigns have put the frame inside the merge field, and then every lead carries the same eight opening words inside a custom field — which is both a repetition problem and a spintax problem.

2. Output contract

Inputs required per row

The client offer block is the part people forget. It is a fixed list of what the client actually does, written as short capability lines. The model is only allowed to promise something on that list.

Output fields

The anchor and offer item are never sent to anyone. They exist so the guard can check the line mechanically, and so an operator can see why a line came out the way it did.

Coverage expectation

8/10 usable, and 8/10 again on a re-run of the shipped script on the same domains.

Of the 2 failures: one was a grammar break the guard could not see, and one was a company with no description in any source.

⚠️ On the re-run the grammar break was gone and a different row went awkward instead. That is the non-determinism in this playbook, and the rule that follows from it: judge a batch, never a row.

Expect 10 to 20% blanks on a normal ecommerce or SMB list, higher on lists full of tiny companies with thin websites.

Copy-fit rules

• Slots into I think we can help you {{specificity_line}}.
• Starts with a plain verb (track, see, know, build, cut), lowercase, no trailing period.
• Never contains the company name, the words "help you", or the word "specifically". All three are already in the frame.
• 6 to 14 words, one idea, no em dashes.
• Never promises sales, growth, demand, or customers unless the offer block says the client does that.

Reading level is a gate, not an aspiration

Enforce Flesch-Kincaid grade 7 on the generated tail. Above it, one simplify rewrite; still above, the row blanks. Measured: 8 filled tails scored 2.5 to 5.9, none needed the rewrite.

⚠️ Gate the tail, not the rendered sentence — and this is measured, not assumed.

The frame Specifically, I think we can help you. scores 5.7 on its own ("Specifically" is five syllables inside a seven-word stem). Once any 6-to-14-word tail is added, the full sentence lands at 7.0 to 9.8 no matter how plain the tail is. A tail of grade 2.5 still renders at 7.0.

Gating the render would blank every row for a property of house copy, not of the data.

The clean fix is a copy change, not a data change: dropping the word "Specifically" takes the worst sampled render from grade 9.1 to 6.3, under the bar, with no change to the generated tail. Do that, and keep gating the tail anyway — it is the half you generate.

⚠️ If you change the frame, change it in the copy and in the prompt's context sentence together, and re-grade on the same domains. The graded verdict was measured against the longer frame.

Downstream gate

The default posture: a row missing this campaign's personalization signal should generally not be uploaded to this campaign at all. Build the list so the line lands, and route the abstains to a campaign whose copy does not need it. That is also what lets you measure whether the line earns its keep.

Where a genuinely optional clause abstains on a small share of rows, pre-render the whole sentence — frame plus value plus trailing period and space — into one field, so the body carries a single merge field that renders as nothing when empty.

⛔ Spintax is banned as the blank-handling mechanism. It picks a variant at random and cannot branch on whether a variable is empty.

A blank is also a soft targeting signal: if more than 30% of a segment comes back blank, the list is wrong for this campaign, not the prompt.